ComfyUI-MOSS-Merge-Presets Introduction
ComfyUI-MOSS-Merge-Presets is a ComfyUI custom node that blends MOSS-Image-Base and MOSS-Image-SFT into a single image-generation model. You can choose a ready-made preset or adjust individual parts of the merge yourself.
The author’s goal is to retain Base’s composition and natural realism while selectively introducing SFT’s strengths in typography, materials, and fine detail.
This can help when you like the overall arrangement or photographic character of Base, but want stronger lettering, richer surfaces, or a more polished finish. Instead of choosing one model for everything, you can explore a balance between them.
The extension provides:
- Three ready-made presets: Base-preserve, Balanced, and SFT-detail.
- A Custom mode with 51 editable merge components.
- Individual control over all 42 MOSS-Image diffusion-transformer blocks.
- A
strengthcontrol for adjusting the overall influence of a preset. - A
merge_inforeadout showing the effective merge weights. - Support for native ComfyUI MOSS-Image workflows, including int8c / convrot Base + SFT merging.
- Compatibility with ComfyUI’s built-in model-saving and model-only LoRA nodes.
- No extra Python dependencies beyond ComfyUI itself.
This is an unofficial community project. It is not affiliated with OpenMOSS, MOSI, or the ComfyUI project. Model weights are not included.
How ComfyUI-MOSS-Merge-Presets Works
Blending model behavior, not finished images
The extension merges the internal weights of two models before image generation. It does not generate two pictures and blend their pixels.
Think of Base and SFT as two versions of the same set of painting tools. Rather than replacing the whole toolbox, the extension lets you combine different proportions of those tools: a Base-heavy approach to composition, for example, with more SFT influence in parts associated with surface detail.
Each merge component has a weight:
0uses Base for that component.1uses SFT for that component.0.3blends 70% Base with 30% SFT for that component. In simplified form:
text Merged weight = Base weight × (1 − effective merge weight) + SFT weight × effective merge weight
These percentages describe the model weights, not a guaranteed percentage of visible style or detail in the final image.
Why different components use different weights
MOSS-Image contains 42 main diffusion-transformer blocks, alongside other components. The presets use different amounts of SFT influence across these components rather than applying one identical blend everywhere.
For example, the presets generally keep early layers more Base-heavy and allow more SFT influence in later layers. The author’s experiments associate early layers more strongly with broad composition and later layers with materials and fine detail.
These relationships are useful starting points, not isolated switches. Changing a layer can affect more than one visual quality.
What strength changes
strength multiplies every component weight in the selected preset or Custom configuration. Each effective component weight is capped at 1.0.
text Effective component weight = component weight × strength capped at 1.0
| strength | Meaning |
|---|---|
| 0.0 | Uses Base throughout the merge. |
| 0.1 or 0.3 | Applies a reduced amount of the selected blend. |
| 1.0 | Uses the selected preset’s defined weights. |
| Above 1.0 | Increases SFT influence until individual components reach their cap. |
For example, Balanced gives layers.36-41 a weight of 0.72:
- At
strength = 0.5, the effective weight is0.36. - At
strength = 1.0, it is0.72. - At
strength = 1.5, the calculated value is1.08, but the effective value is capped at1.0. Because components reach that cap at different times, higher strength does not always produce a proportional increase across the whole model. SFT-detail at1.25is still a blend, not pure SFT.
Use merge_info to check the weights actually being applied.
ComfyUI-MOSS-Merge-Presets Features
Ready-made presets
The presets offer repeatable starting points without requiring you to adjust dozens of controls.
| Preset | Intended balance | Useful starting point for | |---|---|---| | Base-preserve | Keeps more Base influence while adding selected SFT characteristics. | Portraits or scenes where you want to retain Base’s composition and natural appearance. | | Balanced | Provides an intermediate blend of Base and SFT. | General experimentation and your first comparison. | | SFT-detail | Introduces stronger SFT influence, especially in later layers. | Typography, product surfaces, materials, and fine-detail exploration. |
These are starting points, not guarantees. Results remain sensitive to the prompt, seed, sampling settings, and scene.
Exact preset weights
The table below shows the component weights before strength scaling. A range such as layers.0-7 means each layer in that range uses the listed value.
| Component | Base-preserve | Balanced | SFT-detail |
|---|---:|---:|---:|
| video_embedder | 0.12 | 0.20 | 0.30 |
| text_embedder.proj | 0.12 | 0.20 | 0.30 |
| text_embedder.blocks.0 | 0.10 | 0.16 | 0.24 |
| text_embedder.blocks.1 | 0.10 | 0.16 | 0.24 |
| time_embedding | 0.14 | 0.24 | 0.36 |
| shared_block_modulation | 0.12 | 0.20 | 0.30 |
| layers.0-7 | 0.08 | 0.15 | 0.25 |
| layers.8-13 | 0.12 | 0.20 | 0.32 |
| layers.14-21 | 0.18 | 0.30 | 0.50 |
| layers.22-29 | 0.28 | 0.45 | 0.65 |
| layers.30-35 | 0.42 | 0.60 | 0.80 |
| layers.36-41 | 0.52 | 0.72 | 0.90 |
| final_norm_video | 0.12 | 0.20 | 0.30 |
| final_linear_video | 0.25 | 0.35 | 0.50 |
| video_head_modulation | 0.12 | 0.20 | 0.30 |
Custom mode: 51 editable components
Choose Custom to expose all 51 controls. Custom mode starts from the Balanced values.
| Controls | Components |
|---|---|
| m_video_embedder | Image/video embedding component |
| m_text_proj | Text projection |
| m_text_block0, m_text_block1 | Two text-embedding blocks |
| m_time | Time embedding |
| m_shared | Shared block modulation |
| m_L00 through m_L41 | All 42 main transformer blocks |
| m_final_norm | Final normalization |
| m_linear | Final linear output component |
| m_head | Output-head modulation |
You do not need to understand every component to start experimenting. Change a small group of controls, compare the result, and keep the changes that help your image.
Practical layer guidance
The following guidance comes from the author’s controlled Base/SFT merge experiments. It is not an official claim about MOSS training or a guarantee that each layer controls only one feature.
| Component or layer range | Observed influence |
|---|---|
| text_embedder.* | Prompt meaning and text-conditioning projection |
| time_embedding | Global contrast, lighting, and response across sampling timesteps |
| layers.0-13 | Subject placement, overall composition, and broad layout |
| layers.14-27 | Typography hierarchy, structured layout, and mid-level detail |
| layers.28-41 | Material texture, fine detail, and finish |
| final_* / video_head_modulation | Output mapping, color, sharpness, and final appearance |
To preserve composition: keep early-layer merge weights low.
To add surface detail while limiting composition changes: increase late-layer weights first, then explore the middle layers.
To improve typography: experiment with middle-layer weights and compare the exact wording, grouping, and placement of text.
Effective-weight readout
The merge_info field shows the effective weights after strength scaling. This helps you:
- Check how strongly SFT is being used.
- See which components have reached the
1.0cap. - Inspect preset values while the Custom controls are hidden.
- Record the effective settings used for a comparison.
Native int8c / convrot support
Both base_model and sft_model can use supported int8c / convrot quantized models.
In simple terms, quantization stores model weights in a more compact representation. During merging, ComfyUI’s native quantized patch path:
- Dequantizes SFT weights as needed.
- Performs the component-by-component blend.
- Requantizes the result using Base’s quantization layout.
The output retains the Base model’s int8c layout instead of falling back to BF16. Quantization metadata such as
weight_scaleandcomfy_quantis not blended as ordinary model weights; scales are regenerated during requantization. The first patch application can be slower than a BF16 merge because quantized weights are dequantized on demand.
ComfyUI-MOSS-Merge-Presets Models and Compatibility
The extension does not supply its own image models. It is designed to merge these two MOSS-Image models:
| Model | Description | Role in the merge |
|---|---|---|
| MOSS-Image-Base | The base text-to-image model. | Connect to base_model; provides the starting weights. |
| MOSS-Image-SFT | A version fine-tuned from Base on curated data. | Connect to sft_model; contributes according to the merge weights. |
The author’s presets aim to combine Base’s composition and natural realism with SFT’s typography, material, and fine-detail strengths. Upstream examples also demonstrate clearer text grouping and more organized layouts with SFT.
Base-preserve, Balanced, and SFT-detail are merge configurations—not separate downloadable models.
Workflow requirements
For the documented workflow, you need:
- ComfyUI
0.42.0or newer, as recommended by the author. - MOSS-Image Base and SFT diffusion-model weights.
- A ComfyUI-compatible MOSS-Image text encoder.
- A Flux.2 VAE.
- Native ComfyUI MOSS-Image loading and sampling support. ComfyUI-format MOSS-Image weights are available upstream.
The author tested the extension with ComfyUI 0.42.0 on Windows and an RTX 5090. This describes the reported test setup, not a minimum hardware requirement.
Using ComfyUI-MOSS-Merge-Presets
Start with the example workflow
Download moss-merge-model.json, then drag the JSON file onto the ComfyUI canvas.
The example demonstrates:
- Loading Base and SFT diffusion models.
- Merging them with Balanced and
strength = 1.0. - Optionally expanding the prompt with MOSS Prompt Enhancer.
- Sampling the merged model.
- Saving the generated image.
The example includes
MOSSPromptEnhancerLocalfrom the separate ComfyUI-MOSS-PromptEnhancer project. If you do not use prompt expansion, remove or bypass that node and route your positive prompt through the workflow’s normal text-conditioning path.
Reselect the Base, SFT, text encoder, and VAE files in their loader nodes if your local filenames differ.
Connect the merge node
- Add two diffusion-model loaders: one for Base and one for SFT.
- Find MOSS Base+SFT Merge Preset under
model/merging/moss. - Double-click the node-library search result to add it. In some current frontend versions, a single click only previews the node.
- Connect the Base loader’s
MODELoutput tobase_model. - Connect the SFT loader’s
MODELoutput tosft_model. - Connect the merged
MODELoutput to your sampling chain. - Start with Balanced and
strength = 1.0. text Base MODEL ─┐ ├─ MOSS Base+SFT Merge Preset SFT MODEL ──┘ │ ▼ ModelSamplingSD3 │ ▼ SkipLayerGuidanceDiTSimple │ ▼ KSampler
This diagram shows the model path only. Your complete workflow still needs prompt conditioning, latent/image dimensions, VAE decoding, and image output.
Recommended sampling settings
The author tested the presets with this native ComfyUI MOSS-Image setup:
| Setting | Recommended starting value |
|---|---|
| ModelSamplingSD3 shift | 5.0 |
| CFG | 3.5 |
| Sampler | er_sde |
| Scheduler | simple |
| Draft steps | 24 |
| Final-image steps | 40–50 |
| SkipLayerGuidanceDiTSimple.single_layers | 22 |
| SkipLayerGuidanceDiTSimple.start_percent | 0.0 |
| SkipLayerGuidanceDiTSimple.end_percent | 0.5 |
The official Python/Diffusers pipeline uses a separate default setup: 50 steps, Euler, CFG 3.0, shift 5, and negative-branch layer skipping at layer 22. These are upstream pipeline defaults, not identical settings for the ComfyUI example.
Compare presets fairly
Use the test prompts for portraits, products, typography, interiors, anime, and materials.
For a useful comparison:
- Keep the prompt, seed, dimensions, sampler, scheduler, CFG, and step count unchanged.
- Generate each preset at the same
strength. - Compare composition, skin or material texture, text accuracy, and fine detail.
- Repeat with at least two or three seeds.
- Change one setting or layer group at a time when refining a Custom merge. The author provides this common negative prompt:
text nsfw, worst quality, bad anatomy, watermark, jpeg artifact
Visual comparisons
The material and typography sheets show, from left to right: Base, SFT, Base-preserve, Balanced, and SFT-detail. The portrait sheet also includes Balanced at strengths 0.1 and 0.3.
Portrait comparison
Portrait comparison of Base, SFT, and merge presets, including reduced Balanced strengths
Material and realism comparison
Teapot comparison showing differences in material appearance and realism
Typography and graphic-detail comparison
Poster comparison showing typography and graphic-detail differences
Saving Models and Using LoRA with ComfyUI-MOSS-Merge-Presets
Save a merged diffusion model
Use ComfyUI’s built-in ModelSave node, shown as Save Model in some locales, under model/merging.
-
Connect the merge node’s
MODELoutput to ModelSave’smodelinput. -
Set
filename_prefixwithout a file extension, for example: text diffusion_models/moss_balanced_s100 -
Run the workflow. For that prefix, ComfyUI saves the file under:
text ComfyUI/output/diffusion_models/
ComfyUI adds its normal counter-based filename numbering.
To load the saved .safetensors file later with Load Diffusion Model, move it into:
text ComfyUI/models/diffusion_models/
You can also use the equivalent configured model directory.
ModelSave is an output node and has no
MODELoutput. Branch it from the merge node rather than placing it in the middle of the sampling chain.
text Base MODEL ─┐ ├─ Merge Preset ─┬─ ModelSave SFT MODEL ──┘ └─ Sampling chain
The saved file contains the merged diffusion model only. It does not include the text encoder or VAE.
Apply a LoRA after merging
Use ComfyUI’s model-only Load LoRA node, LoraLoaderModelOnly, under model/loaders.
- Connect the merge output to the LoRA node’s
modelinput. - Select the LoRA file.
- Adjust
strength_model. - Connect the LoRA node’s model output to the sampling chain. text Base MODEL ─┐ ├─ Merge Preset ─ Load LoRA ─ Sampling chain SFT MODEL ──┘
Multiple model-only LoRA nodes can be chained. The LoRA is applied to the merged model at inference time.
If you want to save the Base+SFT blend independently of later LoRA adjustments, branch ModelSave directly from the merge output.
MOSS Base+SFT Merge Preset connected to a model-only Load LoRA node
What's New with ComfyUI-MOSS-Merge-Presets
The author maintains the extension’s release history in CHANGELOG.md.
The documented feature set includes ready-made presets, 51-component Custom control, effective-weight reporting, native int8c / convrot merging, and workflows for saving merges and applying LoRAs.
For version-specific changes, consult the changelog. No dated or numbered extension release entries are included here because they were not supplied in the available documentation.
Troubleshooting ComfyUI-MOSS-Merge-Presets
The node appears in search, but only a preview opens
- Locate MOSS Base+SFT Merge Preset in the node library.
- Double-click the result.
- Check the canvas for the added node. Some ComfyUI frontend versions use a single click for preview rather than insertion.
The Custom controls disappear when I choose a preset
This is expected.
- Read
merge_infoto inspect the preset’s effective weights. - Select Custom when you want editable controls.
- Remember that Custom mode starts from Balanced values.
The result changes too much toward SFT
- Lower
strength. - Try Base-preserve.
- In Custom mode, reduce
layers.0-13andtime_embeddingfirst. - Keep the rest of the generation settings fixed while comparing.
- Test another seed to check whether the composition change is seed-sensitive.
Typography or fine detail is still weak
- Try SFT-detail as your starting point.
- In Custom mode, increase
layers.14-29. - Then explore increasing
layers.30-41. - Check
merge_infoto see the effective values. - Compare both the improvement and any unwanted changes to composition or appearance. A merge does not guarantee correct lettering in every image.
The example workflow contains a missing Prompt Enhancer node
- Check whether the missing node is
MOSSPromptEnhancerLocal. - If you want prompt expansion, use the separate MOSS Prompt Enhancer project.
- Otherwise, remove or bypass that node.
- Route your positive prompt through the workflow’s normal text-conditioning path. The Prompt Enhancer is optional for merging.
The example cannot find my model files
- Open the Base loader and choose your local Base model.
- Open the SFT loader and choose your local SFT model.
- Reselect the text encoder and VAE.
- Confirm that Base connects to
base_modeland SFT connects tosft_model. The example’s filenames may differ from yours.
The first quantized merge is slower than expected
For supported int8c / convrot models, the first patch application can take longer than a BF16 merge because dequantization happens on demand.
Check that you are using the documented quantized Base/SFT models and a compatible ComfyUI version. An initial delay alone does not indicate that the output has fallen back to BF16. ### ModelSave will not connect onward to the sampler ModelSave has no model output.
- Disconnect it from the middle of the model path.
- Connect the merge output directly to your sampling chain.
- Create a separate branch from that same merge output to ModelSave.
Frequently asked questions
Does this extension generate images by itself?
No. It outputs a merged MODEL for your normal ComfyUI image-generation workflow.
Is this a training tool? No. It interpolates existing model weights; it does not fine-tune the models.
Does strength = 1.0 mean pure SFT?
No. It applies the selected preset’s weights. Those weights are mostly below 1.0.
Can I preserve composition perfectly while changing only texture? Not reliably. Linear weight merging does not isolate visual properties perfectly, and composition can remain seed-sensitive.
Are all model formats and community variants supported? The documentation specifically covers MOSS-Image Base and SFT, including native int8c / convrot merging. It does not establish compatibility with every community variant or quantization format.
Do I need a separate save extension? No. Use ComfyUI’s built-in ModelSave node.
Does a saved merge include my text encoder and VAE? No. You still need to load those separately.
ComfyUI-MOSS-Merge-Presets Limitations
- Layer effects overlap. Linear weight merging is not equivalent to independent layer-activation or feature control.
- Composition can change between seeds and blends. Test more than one seed before choosing a final configuration.
- Layer guidance is empirical. SFT training details are not fully public, so the guidance reflects the author’s experiments rather than a complete explanation of the model.
- Higher strength is not always better. Components can reach their cap at different times, and stronger SFT influence can alter qualities you wanted to preserve.
- Model weights are separate. The extension does not include Base, SFT, the text encoder, or the VAE.
Learn More about ComfyUI-MOSS-Merge-Presets
Extension resources
- Extension repository — project documentation and source files.
- Example merge workflow — a ready-made canvas for exploring the merge.
- Test prompts — repeatable subjects for evaluating presets.
- Changelog — version-specific updates.
- MIT license — license for this extension.
MOSS-Image resources
- MOSS-Image project — upstream documentation and examples.
- English user guide — prompting advice, sampling recommendations, and layout guidance.
- MOSS-Image homepage — an overview of the underlying image model.
- ComfyUI-format model weights — model files prepared for native ComfyUI workflows.
- Official MOSS-Image demo — an upstream way to explore model behavior.
Optional prompt assistance
ComfyUI-MOSS-PromptEnhancer is a separate companion extension that creates prompts rather than images. It can:
- Expand an idea into a detailed descriptive prompt.
- Create structured Layout prompts for text and object placement.
- Analyze a reference image to produce a reusable prompt.
- Use a local GGUF model or a compatible LLM API. Its prompt output can feed the text-conditioning part of your MOSS-Image workflow, while ComfyUI-MOSS-Merge-Presets controls the Base/SFT model blend. For Layout prompts, keep the prompt’s canvas dimensions consistent with your image-generation dimensions.
License
ComfyUI-MOSS-Merge-Presets is licensed under the MIT License. See LICENSE. Separate models, companion extensions, and upstream projects retain their own licenses.
