MOSS Base+SFT Merge Preset:
MOSSMergePreset, displayed in ComfyUI as “MOSS Base+SFT Merge Preset,” combines a MOSS-Image Base model with its supervised fine-tuned, or SFT, counterpart using different blending amounts for different parts of the model. Rather than applying one uniform blend everywhere, it offers ready-made configurations that progressively increase the SFT contribution across groups of layers, along with a custom mode for individually adjusting 51 model components. This gives you a repeatable way to explore combinations that retain more of the Base model or incorporate more of the SFT model without manually configuring every component. The blending direction is consistent throughout: zero retains Base, while one takes SFT for the affected component. The implementation also includes compatibility handling for int8c/convrot checkpoints, including a fallback that excludes quantization metadata from the weights being merged. Preset names describe their relative blending emphasis, not guaranteed visual outcomes; the actual results depend on the checkpoints and your generation workflow.
MOSS Base+SFT Merge Preset Input Parameters:
base_model
The required MODEL connection supplying the MOSS-Image Base model. The node clones this model and applies the selected SFT contributions to that clone, rather than directly modifying the connected Base model. Components with an effective blending weight of 0 retain their Base contribution. Connect the Base checkpoint here, not the SFT checkpoint, because reversing their positions reverses the practical meaning of the controls. There is no numeric range or default model.
sft_model
The required MODEL connection supplying the MOSS-Image SFT model whose diffusion-model weights contribute to the blend. SFT means supervised fine-tuning: this is the fine-tuned counterpart you want to mix with Base. The node extracts patches from this model under the diffusion_model. prefix and applies them according to the component weights. Use a corresponding MOSS-Image model with compatible parameter names and dimensions; the supplied code does not establish compatibility between arbitrary checkpoints. There is no numeric range or default model.
preset
The required selection controlling where the component blending weights come from. The available options, in order, are Base-preserve, Balanced, SFT-detail, and Custom. The first three use fixed component-weight tables; Custom uses all 51 individual component controls instead. Base-preserve assigns the lowest SFT contributions, Balanced assigns intermediate contributions, and SFT-detail assigns the highest contributions among the predefined choices. The source does not declare a separate default value for this selection; Base-preserve is the first listed option. Individual custom values do not affect execution while a predefined preset is selected.
strength
The required FLOAT multiplier applied to every selected component weight before merging. Its minimum is 0.0, maximum is 1.25, default is 1.0, and adjustment step is 0.01. The effective SFT contribution is the component weight multiplied by strength, then limited to the range 0–1; the Base contribution is one minus that result. A value of 0 retains Base, 1 applies the selected weights unchanged, and 1.25 increases those weights by 25% without allowing any component to exceed full SFT contribution. For example, a component weight of 0.60 becomes 0.30 at strength 0.50 or 0.75 at strength 1.25.
merge_info
The required STRING field intended as a read-only, multiline display of effective component weights after strength scaling. Its default is an empty string, and no length limit or list of options is specified. The merge method accepts this field but never reads its contents, so changing its text does not change the model blend. Treat it as informational rather than as a place to enter instructions or edit weights; the supplied Python class does not itself generate the displayed text.
m_video_embedder
The required FLOAT controlling the SFT contribution for parameters beginning with video_embedder. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.20, and step is 0.01. Zero retains Base for this component, one selects SFT before global strength scaling, and intermediate values blend the two. The prefix is an internal MOSS-Image component name and does not indicate that this node produces video. This control is ignored when a predefined preset is selected.
m_text_proj
The required FLOAT controlling the SFT contribution for text_embedder.proj. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.20, and step is 0.01. It adjusts the matching text-embedding projection component of the diffusion model, not a separately connected text encoder. Zero retains Base, one selects SFT before strength scaling, and intermediate values blend their contributions. Predefined presets use their own value instead.
m_text_block0
The required FLOAT controlling text_embedder.blocks.0., the first of the two listed text-embedding blocks, in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.16, and step is 0.01. It sets the SFT contribution for matching parameters before global strength scaling: 0 retains Base and 1 selects SFT. The supplied implementation does not establish a specific visual effect for this block, so evaluate adjustments with controlled image comparisons. Predefined presets ignore this field.
m_text_block1
The required FLOAT controlling text_embedder.blocks.1., the second listed text-embedding block, in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.16, and step is 0.01. Higher values increase the SFT contribution to this block, while lower values preserve more Base; strength further scales the selected amount. This field does not change the first text block and is ignored when using a predefined preset.
m_time
The required FLOAT controlling the SFT contribution for time_embedding. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.24, and step is 0.01. It blends the model’s time-embedding component, which participates in representing the diffusion process’s timestep; it does not change sampling steps or a timeline duration. Zero retains Base and one selects SFT before strength scaling. Predefined presets replace this custom value with their own setting.
m_shared
The required FLOAT controlling shared_block_modulation. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.20, and step is 0.01. It sets the blend for the component identified by that shared-modulation prefix, rather than acting as a second global strength control. Zero retains Base, one selects SFT before strength scaling, and intermediate values mix their contributions. This field is ignored for predefined presets.
m_L00
The required FLOAT setting the SFT contribution for layers.0. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.15, and step is 0.01. It affects only parameters matching this layer prefix: 0 retains Base, 1 selects SFT before strength scaling, and intermediate values blend them. The supplied code does not assign this layer a guaranteed visual role. Predefined presets ignore this field.
m_L01
The required FLOAT setting the SFT contribution for layers.1. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.15, and step is 0.01. Lower values preserve more Base in this layer; higher values incorporate more SFT, with the final amount multiplied by strength and clamped to 0–1. It does not control adjacent layers and is ignored under predefined presets.
m_L02
The required FLOAT setting the SFT contribution for layers.2. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.15, and step is 0.01. Zero retains this layer’s Base contribution, while one selects its SFT contribution before global strength scaling. Intermediate settings blend the matching parameters. Predefined presets use their fixed layer weight instead.
m_L03
The required FLOAT setting the SFT contribution for layers.3. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.15, and step is 0.01. Adjust it to change the Base/SFT balance specifically for this layer, remembering that strength scales the chosen value. A value of 0 retains Base and 1 selects SFT before that scaling. It has no effect with predefined presets.
m_L04
The required FLOAT setting the SFT contribution for layers.4. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.15, and step is 0.01. It blends matching layer parameters from Base and SFT, using more SFT as the value increases. The effective contribution also depends on strength and cannot exceed 1. Predefined presets ignore the custom setting.
m_L05
The required FLOAT setting the SFT contribution for layers.5. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.15, and step is 0.01. Zero preserves Base for the matching parameters, one selects SFT before strength scaling, and intermediate values combine their contributions. This adjustment applies only to layer 5 and is ignored when a predefined preset supplies the weights.
m_L06
The required FLOAT setting the SFT contribution for layers.6. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.15, and step is 0.01. It changes the blend for this layer without directly changing other component controls. The final SFT amount is this value multiplied by strength and limited to 0–1. Predefined presets do not use this field.
m_L07
The required FLOAT setting the SFT contribution for layers.7. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.15, and step is 0.01. This is the last layer in the presets’ 0–7 weight group, although Custom mode allows you to adjust it independently. Zero retains Base and one selects SFT before strength scaling. Predefined presets ignore this custom value.
m_L08
The required FLOAT setting the SFT contribution for layers.8. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.20, and step is 0.01. This is the first layer in the presets’ 8–13 group, but its custom setting affects only layer 8. Zero retains Base and one selects SFT before the global multiplier is applied. Predefined presets use their own weight instead.
m_L09
The required FLOAT setting the SFT contribution for layers.9. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.20, and step is 0.01. Increasing it shifts this layer toward SFT, while decreasing it preserves more Base. The effective value is scaled by strength and clamped to 0–1. This control is ignored for predefined presets.
m_L10
The required FLOAT setting the SFT contribution for layers.10. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.20, and step is 0.01. It blends only the matching layer parameters, with 0 retaining Base and 1 selecting SFT before global strength scaling. Its name uses the model’s layer index, not a percentage. Predefined presets ignore this field.
m_L11
The required FLOAT setting the SFT contribution for layers.11. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.20, and step is 0.01. Intermediate values mix this layer’s Base and SFT contributions, while 0 and 1 represent their endpoints before strength scaling. It does not automatically adjust neighboring layers and has no effect under predefined presets.
m_L12
The required FLOAT setting the SFT contribution for layers.12. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.20, and step is 0.01. Use it to independently control how much SFT contributes to this layer. Global strength scales the amount, and the effective result remains within 0–1. Predefined presets replace this setting with their fixed weight.
m_L13
The required FLOAT setting the SFT contribution for layers.13. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.20, and step is 0.01. It is the last layer in the presets’ 8–13 group but remains individually adjustable in Custom mode. Zero retains Base and one selects SFT before strength scaling. Predefined presets ignore the field.
m_L14
The required FLOAT setting the SFT contribution for layers.14. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.30, and step is 0.01. This is the first layer in the presets’ 14–21 group. Its custom value changes only this layer’s blend, from Base at 0 to SFT at 1 before strength scaling. It is ignored for predefined presets.
m_L15
The required FLOAT setting the SFT contribution for layers.15. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.30, and step is 0.01. Higher settings add more of this layer’s SFT parameters, while lower settings retain more Base. The node multiplies the setting by strength and clamps the result to 0–1. Predefined presets do not use this field.
m_L16
The required FLOAT setting the SFT contribution for layers.16. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.30, and step is 0.01. It controls this layer independently, with 0 retaining Base, 1 selecting SFT before strength scaling, and intermediate values blending them. The code does not establish a specific visual feature controlled by this layer. Predefined presets ignore it.
m_L17
The required FLOAT setting the SFT contribution for layers.17. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.30, and step is 0.01. Adjusting this value changes the mix of matching Base and SFT layer parameters. Its effective SFT amount is also governed by strength and cannot exceed 1. It has no effect while a predefined preset is selected.
m_L18
The required FLOAT setting the SFT contribution for layers.18. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.30, and step is 0.01. Zero retains the Base contribution for this layer, while one selects SFT before applying the global multiplier. Intermediate values allow a partial blend. Predefined presets supply their own layer weight and ignore this setting.
m_L19
The required FLOAT setting the SFT contribution for layers.19. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.30, and step is 0.01. It affects only parameters with the corresponding layer prefix, not every layer in the same preset group. Global strength scales its Base/SFT balance. The control is ignored when a predefined preset is active.
m_L20
The required FLOAT setting the SFT contribution for layers.20. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.30, and step is 0.01. Lower values keep this layer closer to Base; higher values incorporate more SFT. Zero and one are the Base and SFT endpoints before strength scaling. Predefined presets do not read this custom value.
m_L21
The required FLOAT setting the SFT contribution for layers.21. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.30, and step is 0.01. This is the final layer in the presets’ 14–21 group, but Custom mode controls it separately. Its effective SFT contribution is the value multiplied by strength and clamped to 0–1. Predefined presets ignore this field.
m_L22
The required FLOAT setting the SFT contribution for layers.22. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.45, and step is 0.01. This starts the presets’ 22–29 layer group, though this control affects only layer 22. Zero retains Base and one selects SFT before global strength scaling. It is ignored when a predefined preset is selected.
m_L23
The required FLOAT setting the SFT contribution for layers.23. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.45, and step is 0.01. It lets you blend this layer independently between Base and SFT. The final amount is scaled by strength and limited to 0–1, so the field alone does not determine the effective contribution. Predefined presets ignore it.
m_L24
The required FLOAT setting the SFT contribution for layers.24. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.45, and step is 0.01. Increasing it increases SFT’s contribution to matching parameters; decreasing it preserves more Base. A value of 1 reaches full SFT only when the scaled value reaches 1. Predefined presets use their own setting instead.
m_L25
The required FLOAT setting the SFT contribution for layers.25. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.45, and step is 0.01. Zero retains Base for this layer, one selects SFT before strength scaling, and intermediate values blend their contributions. It does not change the custom controls for neighboring layers and is ignored under predefined presets.
m_L26
The required FLOAT setting the SFT contribution for layers.26. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.45, and step is 0.01. This determines the initial blend amount for matching layer parameters; strength then scales that amount and the node clamps it to 0–1. Predefined presets replace this field with their fixed layer weight.
m_L27
The required FLOAT setting the SFT contribution for layers.27. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.45, and step is 0.01. It controls this layer’s Base/SFT mixture without assigning a guaranteed visual effect to the adjustment. Zero retains Base and one selects SFT before strength scaling. This field is ignored when using a predefined preset.
m_L28
The required FLOAT setting the SFT contribution for layers.28. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.45, and step is 0.01. Lower values preserve more of the Base parameters in this layer, while higher values incorporate more SFT. Its effective contribution also depends on global strength. Predefined presets do not use this custom setting.
m_L29
The required FLOAT setting the SFT contribution for layers.29. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.45, and step is 0.01. This is the last layer in the presets’ 22–29 group, but you can adjust it independently in Custom mode. Zero retains Base and one selects SFT before strength scaling. Predefined presets ignore this field.
m_L30
The required FLOAT setting the SFT contribution for layers.30. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.60, and step is 0.01. This starts the presets’ 30–35 group, although its custom setting affects only layer 30. The selected SFT amount is multiplied by strength and limited to 0–1. Predefined presets use their fixed weight instead.
m_L31
The required FLOAT setting the SFT contribution for layers.31. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.60, and step is 0.01. Zero retains this layer’s Base contribution, one selects SFT before strength scaling, and intermediate values blend the matching parameters. It does not automatically change other layers and is ignored under predefined presets.
m_L32
The required FLOAT setting the SFT contribution for layers.32. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.60, and step is 0.01. Increasing the value shifts this layer toward SFT; decreasing it keeps more Base. The final blend also depends on strength, with the effective SFT contribution capped at 1. Predefined presets ignore the field.
m_L33
The required FLOAT setting the SFT contribution for layers.33. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.60, and step is 0.01. This independently blends parameters belonging to layer 33, using 0 for Base and 1 for SFT before global strength scaling. The implementation does not identify a specific visual feature controlled by this layer. Predefined presets use their own weight.
m_L34
The required FLOAT setting the SFT contribution for layers.34. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.60, and step is 0.01. It determines how much SFT contributes to this layer before multiplication by strength. Intermediate values retain contributions from both models, while the final effective value remains within 0–1. The field is ignored for predefined presets.
m_L35
The required FLOAT setting the SFT contribution for layers.35. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.60, and step is 0.01. This is the last layer in the presets’ 30–35 group and can be adjusted independently in Custom mode. Zero retains Base and one selects SFT before strength scaling. Predefined presets ignore this custom value.
m_L36
The required FLOAT setting the SFT contribution for layers.36. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.72, and step is 0.01. This begins the presets’ final 36–41 layer group, but the field controls only layer 36. Global strength scales the selected SFT contribution and the node clamps it to 0–1. Predefined presets do not use this field.
m_L37
The required FLOAT setting the SFT contribution for layers.37. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.72, and step is 0.01. It shifts the matching parameters between Base at 0 and SFT at 1 before strength scaling. Adjusting this layer does not directly change the other custom values. Predefined presets use their own fixed weight.
m_L38
The required FLOAT setting the SFT contribution for layers.38. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.72, and step is 0.01. Higher values increase SFT’s contribution to this layer, while lower values retain more Base. The effective amount is multiplied by strength and cannot exceed 1. This setting is ignored while a predefined preset is active.
m_L39
The required FLOAT setting the SFT contribution for layers.39. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.72, and step is 0.01. It controls the blend for this layer only, with intermediate values mixing Base and SFT parameters before global strength scaling. The supplied code does not guarantee a particular visual effect from adjusting it. Predefined presets ignore this field.
m_L40
The required FLOAT setting the SFT contribution for layers.40. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.72, and step is 0.01. Zero retains Base for matching parameters, while one selects SFT before applying strength. The effective contribution is clamped to 0–1 after scaling. Predefined presets substitute their own layer weight.
m_L41
The required FLOAT setting the SFT contribution for layers.41., the last of the 42 indexed layers exposed by this node, in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.72, and step is 0.01. It affects this layer rather than the separate final normalization, final linear, or head-modulation components. Zero retains Base and one selects SFT before strength scaling. Predefined presets ignore this field.
m_final_norm
The required FLOAT controlling final_norm_video. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.20, and step is 0.01. It sets the blend for the final normalization component, independently of the final linear component and indexed layers. Zero retains Base and one selects SFT before global strength scaling. The internal video naming does not change the node’s MODEL output into video, and predefined presets ignore this custom setting.
m_final_linear
The required FLOAT controlling final_linear_video. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.35, and step is 0.01. It adjusts the Base/SFT mixture for the final linear component, separately from final normalization and head modulation. Zero retains Base, one selects SFT before strength scaling, and intermediate values combine their contributions. Predefined presets use their own final-linear weight rather than this field.
m_head
The required FLOAT controlling video_head_modulation. in Custom mode. Its minimum is 0.0, maximum is 1.0, default is 0.20, and step is 0.01. It sets the SFT contribution for the head-modulation component, not the entire model or all final components together. Zero retains Base and one selects SFT before global strength scaling. The supplied code does not establish a specific visual effect for this adjustment, and predefined presets ignore it.
MOSS Base+SFT Merge Preset Output Parameters:
MODEL
The single MODEL output is the cloned Base model with SFT patches applied according to the effective component weights. For each matching parameter, the node requests an SFT contribution equal to the scaled, clamped weight and a Base contribution equal to one minus that weight. Thus, an effective value of 0.25 represents a 25% SFT and 75% Base parameter blend, not a guarantee that an image will exhibit those percentages of either model’s visual characteristics. Parameters outside the selected component prefixes receive no SFT contribution, and the merge operates on diffusion-model patches rather than separately merging text encoders or VAEs. Connect this output wherever your workflow expects the diffusion MODEL, such as the model connection used for sampling. The output is a model object within the workflow, not an image or an automatically saved checkpoint.
MOSS Base+SFT Merge Preset Usage Tips:
-
Connect the MOSS-Image Base model to
base_modeland its SFT counterpart tosft_model. Keep the checkpoints compatible in architecture and parameter dimensions; this node is not a general-purpose converter between unrelated models. -
Start with
Balancedatstrength1.0 for an intermediate preset blend, then compare it againstBase-preserveandSFT-detailusing the same prompt, seed, sampler, resolution, and other generation settings. -
Use
Base-preservewhen you want the smallest preset SFT contributions. At strength 1.0, its layer groups 0–7, 8–13, 14–21, 22–29, 30–35, and 36–41 use 0.08, 0.12, 0.18, 0.28, 0.42, and 0.52 respectively; its general component weight is 0.12, text-block weights are 0.10, time-embedding weight is 0.14, and final-linear weight is 0.25. - UseBalancedfor intermediate preset weights. At strength 1.0, the same six layer groups use 0.15, 0.20, 0.30, 0.45, 0.60, and 0.72; its general component weight is 0.20, text-block weights are 0.16, time-embedding weight is 0.24, and final-linear weight is 0.35. -
Use
SFT-detailto apply the strongest predefined SFT blend, without assuming it will always produce more detail. At strength 1.0, the six layer groups use 0.25, 0.32, 0.50, 0.65, 0.80, and 0.90; its general component weight is 0.30, text-block weights are 0.24, time-embedding weight is 0.36, and final-linear weight is 0.50. -
In the preset descriptions, the general component weight applies to
video_embedder.,text_embedder.proj.,shared_block_modulation.,final_norm_video., andvideo_head_modulation.. The text blocks, time embedding, and final linear component use the separately stated values. -
Adjust
strengthbefore editing many individual components. For example,Balancedat 0.50 halves every Balanced SFT weight, providing a quick way to reduce the blend while preserving its relative layer pattern. -
Remember that strength above 1.0 saturates some components. With
SFT-detailat 1.25, a weight of 0.90 becomes 1.0 after clamping, while a weight of 0.65 becomes 0.8125; increasing strength does not preserve the original proportions once components hit the limit. -
Select
Custombefore expecting individual component edits to affect the result. The declared custom defaults matchBalanced, but existing workflow values may differ, so inspect them rather than assuming a preset switch has reset them. -
In Custom mode, change one component or a small group at a time and keep comparison settings fixed. The code defines which parameters each control blends, but it does not establish that a particular layer exclusively controls anatomy, texture, composition, or another visual feature.
-
Set
strengthto 0 for a Base-preserving reference. To request full SFT contribution across all 51 exposed components, useCustom, set every component to 1.0, and set strength to 1.0; this still does not merge components outside the node’s supported diffusion-model prefixes. -
For int8c/convrot checkpoints, keep both ComfyUI and the extension current. The fallback patch extraction skips quantization metadata rather than treating scale information as ordinary model weights; the model still needs to be loadable by your ComfyUI installation.
MOSS Base+SFT Merge Preset Common Errors and Solutions:
Required input is missing
- Explanation: ComfyUI may reject an incomplete workflow before the merge runs. Every listed input is declared required, including the 51 custom component fields even when a predefined preset means their values are not used.
- Solution: Connect both MODEL inputs and restore missing widgets or values by recreating the node if necessary. For API workflows, include every required field using the exact interface names.
Value not in list: preset
- Explanation: A workflow or API request may contain a preset name that is not one of the four supported options. If validation is bypassed, the direct preset lookup can instead raise a
KeyErrorfor the invalid name. - Solution: Use exactly
Base-preserve,Balanced,SFT-detail, orCustom, including capitalization and hyphens. Reselect the preset if loading an older workflow.
Value outside the allowed range
- Explanation: ComfyUI validation may reject a
strengthoutside 0.0–1.25 or a custom component value outside 0.0–1.0. Exact validation wording depends on the ComfyUI version. - Solution: Keep strength within 0.0–1.25 and all 51 component weights within 0.0–1.0. The internal clamp protects effective blending ratios, but it does not replace input validation.
KeyError for a custom field, such as 'm_L00'
- Explanation: In Custom mode, the merge method directly reads every custom field. A direct call or incomplete API submission that omits one can raise a
KeyErrornaming the missing field. - Solution: Supply all 51 component values, including non-layer controls such as
m_timeandm_head. Recreating the node in ComfyUI restores the declared widget defaults.
ValueError: could not convert string to float
- Explanation: Custom mode converts each component value to a number. A direct call that supplies text such as a percentage label or an instruction instead of a numeric value can fail during conversion.
- Solution: Use decimal numbers such as 0.25 rather than text such as “25%.” Keep annotations in your workflow notes, not in numeric component fields.
AttributeError involving clone or get_key_patches
- Explanation: These methods belong to the model objects the node expects. An invalid object supplied through a direct call, or an incompatible model wrapper, may not expose them. An
AttributeErrorduring native SFT patch extraction is normally caught and triggers the compatibility fallback, but failures elsewhere may still stop execution. - Solution: Use MODEL outputs from compatible checkpoint loaders and verify that each checkpoint works in a basic generation workflow. If the error persists, update ComfyUI and the extension, then inspect the complete traceback to identify which object or method failed.
MOSS merge: using quantized-model-compatible patch extraction (...)
- Explanation: This is an informational log message present in the source, not automatically an error. It means native patch extraction raised an
AttributeErrorand the node is trying its fallback for compatible quantized models. - Solution: If the workflow completes, no corrective action is needed. If it fails afterward, troubleshoot the subsequent exception rather than treating this informational message as the failure itself.
Tensor size or shape mismatch during merging or sampling
- Explanation: Compatible names alone do not guarantee compatible parameter dimensions. Different architectures or incompatible checkpoint variants may cause tensor-size errors when patches are applied or when the merged model is used; the exact message comes from ComfyUI or PyTorch.
- Solution: Check that Base and SFT belong to the intended compatible MOSS-Image architecture. Test them separately and avoid combining unrelated models merely because both loaders produce MODEL connections.
CUDA out of memory
- Explanation: Depending on your hardware and ComfyUI’s loading behavior, holding the source models and applying merged patches may exceed available GPU memory. This is an environment-level error rather than a custom error raised by the node.
- Solution: Close other GPU-heavy applications, unload unrelated models, and use ComfyUI’s appropriate memory-management settings. If the error occurs during sampling, reduce resolution or batch size; lowering
strengthalone does not reduce the number of model parameters.
Custom adjustments have no visible effect
- Explanation: This is a troubleshooting symptom rather than a defined exception. Predefined presets ignore custom values, strength 0 removes all SFT contribution, and a downstream sampler may still be connected to the original model. Similar source checkpoints can also make changes difficult to notice.
- Solution: Select
Custom, use a nonzero strength, and confirm that sampling uses this node’s MODEL output. Compare a controlled test at component values 0 and 1 while keeping the prompt, seed, and sampling settings fixed.
Editing merge_info does not change the result
- Explanation: The merge method accepts
merge_infobut does not use its contents. It is intended as an informational display, not an editable merge configuration. - Solution: Change
preset,strength, or the individual component controls in Custom mode. If the display appears stale, check the effective settings directly and verify that the extension’s frontend assets loaded correctly.
Selecting SFT-detail does not produce a pure SFT model
- Explanation:
SFT-detailis a layered blend, not a full-SFT switch. At strength 1.0, its component weights range from 0.24 to 0.90, so Base still contributes throughout the configured components. - Solution: For full SFT contribution across the supported components, select
Custom, set all 51 component weights to 1.0, and use strength 1.0. Do not interpret this as replacing separately loaded text encoders, VAEs, or unsupported model components.
The merged model produces poor or unexpected images
- Explanation: Successful parameter blending does not guarantee improved image quality. Swapped source connections, incompatible checkpoints, or a blend that does not suit your prompts can produce unexpected results, and preset names do not guarantee particular visual improvements.
- Solution: Verify the Base/SFT connection order, test each source model independently, and compare the three presets using fixed generation settings. Reduce strength or return to the Balanced custom defaults before making small, isolated adjustments.
