ComfyUI-Sigmax Introduction
ComfyUI-Sigmax creates sigma schedules for supported image and video models in ComfyUI. A sigma schedule tells a sampler how the amount of noise should change over the steps of a generation. Different models expect different schedules, so using a schedule designed for the wrong model or variant can produce an unintended workflow.
Sigmax gives you model-specific nodes with explicit choices for variants, source recipes, steps, dimensions, and other relevant settings. It also gives you tools to inspect, compare, and edit schedules. The extension registers 25 namespaced nodes; search for Sigmax in ComfyUI’s node menu to find them.
What Sigmax does not do: it does not download or load model weights, replace ComfyUI’s sampler, apply a LoRA, or guarantee image or video quality. Its main output is SIGMAS, which you connect to a custom-sampling workflow that accepts external sigmas. You still need the matching model and the other components required by that workflow.
How ComfyUI-Sigmax Works
Think of generation as gradually revealing an image or video from noise. The sampler performs the steps; the sigma schedule is the sequence of noise levels it follows. Sigmax prepares that sequence using a recipe associated with the model and mode you select.
A typical workflow is:
- Choose a Sigmax scheduler for your model family, such as Qwen Image, Wan, or LTX.
- Select the exact variant and recipe where the node offers a choice. For example, Krea 2
RAWandTurbohave different recommended settings. - Enter settings that describe your intended output, such as step count or dimensions, when the recipe requires them.
- Connect the node’s
SIGMASoutput directly to the appropriate custom-sampling path. - Read
schedule_infoto check the selected recipe, effective settings, and any warnings.
Use one external-sigma source for that sampling path. Passing Sigmax’s output through another scheduler, or adding an extra time shift, can change the schedule again. Sigmax constructs a schedule; it does not automatically configure every other node in your workflow.
Some nodes offer strict_official. Keep it enabled when you want to stay within the listed recipe settings. If you deliberately try a different step count or an experimental recipe, review the node’s requirements and the resulting schedule_info.
ComfyUI-Sigmax Features
Model-specific schedule nodes
Sigmax provides separate schedules for supported model families instead of silently falling back to a generic recipe. This makes choices that matter to your workflow visible: a Base model may need different settings from a Turbo or Distilled variant, and a video task may need a specified resolution or stage.
Sigmax.ModelAwareSigmaScheduler can help identify a model family, but it may not be able to determine an ambiguous variant. If it cannot distinguish Krea 2 RAW from Turbo, for example, choose the variant yourself rather than relying on a filename.
Schedule information and comparison
Sigmax.ProfileInspectorshows details about a model profile.Sigmax.ScheduleInspectorhelps you examine a constructed schedule.Sigmax.ScheduleComparisonlets you compare two schedules.- Versioned schedule information and fingerprints help you record which recipe and settings a workflow used.
These tools are useful when you are testing two recipes or returning to an older workflow and want to understand how its schedule was built.
Schedule editing
Sigmax.ScheduleSlice, Sigmax.ScheduleConcatenate, and Sigmax.ScheduleResample let you take part of a schedule, join schedules, or change the number of schedule points. They retain traceable schedule information. These are editing tools for deliberate workflow experiments; an edited schedule is not automatically an official model recipe.
Checkpoint evidence
Sigmax.CheckpointEvidenceInspector reads a bounded header from a local .safetensors file without loading its model weights. It may suggest a variant, but that suggestion is advisory, not proof that the checkpoint matches a particular recipe.
Experimental tools
Sigmax.Krea2ConditioningRebalance is an optional experimental CONDITIONING adjustment for explicitly selected Krea 2 RAW or Turbo workflows. At strength 0, it leaves the conditioning unchanged. It does not patch the model or scheduler.
Sigmax.AdvancedFlowMatchScheduler is an experimental schedule constructor, not a universal profile for arbitrary models and not a replacement sampler. The author also maintains internal sampler research checks, but these are not registered as public sampler nodes.
ComfyUI-Sigmax Models
The settings below are starting points or specified recipe settings from the provided documentation. Match the Sigmax node to the exact model variant you are using, and check schedule_info before sampling.
Image models
| Model and node | When to use it and what to check |
|---|---|
Krea 2 — Sigmax.Krea2SigmaScheduler | Select Turbo or RAW explicitly and enter your intended width and height. The listed Turbo recipe uses 8 steps, Euler, and CFG 1.0. RAW lists either 52 steps at CFG 4.5 or 28 steps at CFG 5.5. Experimental RAW-to-Turbo LoRA modes require a compatible LoRA applied separately to RAW; a community starting point is 12 steps, Euler, and CFG 1.0. |
Z-Image — Sigmax.ZImageSigmaScheduler | Choose Base or Turbo. Base lists 28–50 steps, with 50 as the default and CFG 4.0. Turbo lists 8 steps and CFG 1.0. |
FLUX.1-schnell — Sigmax.Flux1SchnellSigmaScheduler | Use for the schnell family, with 1–4 steps, 4 by default, and CFG 1.0. This does not imply support for other FLUX variants. |
Original Qwen Image — Sigmax.QwenImageSigmaScheduler | Comfy Native follows the documented native schedule. Existing Comfy Fixed workflows keep their older, numerically different behavior; Diffusers Dynamic is another explicit option. The listed compatibility setting is 50 steps with host true CFG 4.0. Choose a mode deliberately when revisiting a saved workflow. |
Qwen Image 2.1 — Sigmax.QwenImage21SigmaScheduler | Use this separate node. Comfy Native lists 25 steps, Euler/Simple, and true CFG 1.0. Diffusers Dynamic lists 40 steps and needs one explicit target: dimensions or target tokens. Changing modes does not change the step count automatically. |
Original Stable Diffusion 3 Medium — Sigmax.SD3SigmaScheduler | Choose the source mode: Publisher Reference (1.0) at 50 steps or Comfy/Diffusers Fixed (3.0) at 28 steps. This profile does not cover SD3.5 or Turbo. |
AuraFlow v0.2 — Sigmax.AuraFlowSigmaScheduler | Use the explicit Official Fixed (1.73) mode for the original fal AuraFlow v0.2 schedule, listed at 50 steps and CFG 3.5. |
Lumina-Image 2.0 — Sigmax.Lumina2SigmaScheduler | Use Official Fixed (6.0) for the original Alpha-VLLM text-to-image schedule, listed at 50 steps and CFG 4.0. |
HunyuanImage 2.1 — Sigmax.HunyuanImage21SigmaScheduler | Select Base (5.0) for the listed 50-step, CFG 3.5 recipe, or Distilled (4.0) for the listed 8-step, CFG 3.25 recipe. The extension constructs the schedule, not the model’s encoders or conditioning. |
Anima Base v1.0, Aesthetic, or Turbo — Sigmax.AnimaSigmaScheduler | Choose the variant explicitly. Base lists 30–50 steps, default 50, at CFG 4.5; Aesthetic uses the Base recipe. Turbo lists 8–12 steps at CFG 1.0. |
For Qwen Image 2.1 native mode, select None (Native) as the target source and leave width, height, and tokens at zero. For dynamic mode, choose either Dimensions with positive width and height divisible by 32, leaving tokens at zero, or Target Tokens with a positive image_seq_len, leaving dimensions at zero. With strict_official enabled, use 25 native or 40 dynamic steps. The extension includes native and dynamic model-free examples, as well as an original Qwen Image native example. These examples construct schedules; they do not demonstrate finished-image quality. Qwen Image 2.1 model use is separately subject to the publisher’s Qwen Research License, including its restrictions on commercial use.
Video and audio-video models
| Model and node | When to use it and what to check |
|---|---|
MiniMax H3 Base FL2VA/Ref2VA and experimental Turbo — Sigmax.MiniMaxH3SigmaScheduler | Select FL2VA or Ref2VA. The default h3_endpoint schedule needs no MODEL input. Nine experimental ComfyUI-native scheduler choices require an already-shifted H3 MODEL with shifts matching the selected recipe. Turbo recipe selection does not load its required LoRA. |
Wan 2.1 and Wan 2.2 — Sigmax.WanSigmaScheduler | Choose generation, task, source, and resolution as applicable. Profiles cover Wan 2.1 T2V, I2V, FLF2V, and VACE, plus Wan 2.2 TI2V, T2V, I2V, S2V, and Animate. For Wan 2.1 I2V, specify 480P or 720P. |
Wan Animate 2 — Sigmax.WanSigmaScheduler | Choose Base, Distilled, or the separate Comfy-optimized lane. Base lists 40 steps; Distilled lists 10. The optimized lane is specifically for its matching model and step-distilled LoRA at 480P, using LCM, Simple sigmas, 6 steps, CFG 1.0, and shift 5.0. It does not apply the LoRA or divide video into chunks for you. |
LTXV 0.9.8, LTX-2 19B, and LTX-2.3 22B — Sigmax.LTXSigmaScheduler | Select the generation and stage explicitly. Dev profiles use a shift derived from token count; distilled Stage 1 and Stage 2 modes use specified publisher vectors. The extension does not load video weights or run encoders. |
MiniMax H3’s scheduler menu contains h3_endpoint, simple, sgm_uniform, karras, exponential, ddim_uniform, beta, normal, linear_quadratic, and kl_optimal. The nine choices other than h3_endpoint are experimental compatibility options, not recommendations about visual quality or performance. Where an explicit H3 shift configuration is needed, use the host’s MiniMaxH3SigmaShift node, also displayed as ModelSamplingMiniMaxH3, after the final model branch and before the guider or Sigmax scheduler. Check the compatibility notes for qualified ComfyUI versions and model-specific requirements.
The optional MiniMax H3 Turbo selector provides these experimental community recipes. You must load the matching LoRA separately:
| Recipe | Steps | Video/audio shift | Task and resolution |
|---|---|---|---|
h3.fl2va.lightx2v-turbo-4-v0.1-544p | 4 | 12.0 / 3.0 | FL2VA, 544p |
h3.fl2va.lightx2v-turbo-8-v1.0-544p | 8 | 12.0 / 3.0 | FL2VA, 544p |
h3.fl2va.lightx2v-turbo-4-v1.0-768p | 4 | 6.0 / 3.0 | FL2VA, 1344×768 |
h3.ref2va.lightx2v-turbo-4-v0.1-544p | 4 | 12.0 / 3.0 | Ref2VA, 544p |
For Wan 2.2 A14B, boundary information is supplied by the calling workflow; Sigmax does not route between model experts. Across the video profiles, schedule construction should not be read as a claim that Sigmax loads weights, handles all conditioning, or establishes video-quality parity.
What’s New with ComfyUI-Sigmax
Version 1.1.3
- Qwen Image choices are clearer. A dedicated Qwen Image 2.1 node adds native and dynamic schedules with explicit dimension or token targets. Original Qwen Image gains an explicit native mode while existing saved modes, defaults, and numerical behavior remain available. This helps you choose a source recipe without silently changing an older workflow.
- Video profiles cover more tasks. Wan support expands to FLF2V, VACE, S2V, Animate, and Wan Animate 2 Base/Distilled. A separate Comfy-optimized Wan Animate 2 recipe records the matching 480P, LCM, Simple, 6-step, CFG 1.0, shift 5.0 setup.
- MiniMax H3 adds experimental choices. The release adds community Turbo recipes and a ten-choice scheduler menu. The node makes model-shift ownership explicit, so you can see when an already-shifted
MODELis required. - Compatibility and validation improve. The author requalified MiniMax H3 native schedulers on ComfyUI 0.39.0 while retaining documented pinned-host support, added Qwen numerical checks, and improved checks for host logs and release artifacts.
- Internal sampler research expands. Deterministic and seeded stochastic Flow Euler checks were extended. These are internal checks; version 1.1.3 does not add a replacement public sampler.
The provided context does not include a complete history of earlier releases.
Troubleshooting ComfyUI-Sigmax
Sigmax nodes do not appear
- Restart ComfyUI.
- Search the node menu for
Sigmax. - Check ComfyUI’s startup log for a loading error.
- Confirm you have only one direct
comfyui-sigmaxfolder underComfyUI/custom_nodes, rather than a duplicate or nested copy.
A node fails to import
- Check that your environment meets the documented minimums: Python 3.10 and ComfyUI 0.29.0.
- Check for duplicate or nested copies of the extension.
- Read the startup error and the compatibility notes, especially for a model-specific feature that may require a qualified host version.
The result looks as though the schedule was shifted twice
- Trace the connection from Sigmax’s
SIGMASoutput to your custom sampler. - Remove any second scheduler applied to those sigmas.
- Check whether another model time shift is also being applied outside the selected recipe. For MiniMax H3 native choices, follow the recipe’s explicit model-shift requirements rather than removing a required H3 model configuration.
Qwen Image 2.1 rejects my settings
- Confirm you are using
Sigmax.QwenImage21SigmaScheduler, not the original Qwen Image node. - With
strict_officialenabled, set 25 steps for native or 40 steps for dynamic. - For native mode, select
None (Native)and set width, height, and tokens to zero. - For dynamic mode, select exactly one target source: valid dimensions or positive target tokens. Leave the inactive controls at zero.
- Recheck the node after switching modes: it does not change
stepsautomatically.
MiniMax H3 asks for MODEL or reports a shift mismatch
- Check the scheduler choice.
h3_endpointis the default pure schedule; the nine experimental native choices require an already-shifted H3MODEL. - Confirm that the selected Base or Turbo recipe matches the model path’s video and audio shifts. Base defaults are
12.0 / 3.0; the listed 768p FL2VA four-step Turbo recipe uses6.0 / 3.0. - If explicit configuration is needed, place
ModelSamplingMiniMaxH3after the final model or LoRA branch and before the guider or scheduler. - Do not add a separate
BasicSchedulerto the path using Sigmax’sSIGMAS.
Frequently asked questions
Does Sigmax download the models or LoRAs I need?
No. It constructs schedules. Obtain and connect any matching model or LoRA separately, subject to that model’s terms.
Does choosing a Turbo recipe make generation faster by itself?
The provided documentation does not claim a speed improvement from the Sigmax selector. Match the specified model, LoRA, task, and workflow settings before evaluating a recipe for your own use.
Can I use one Sigmax schedule for any model?
No. Select a supported family and the appropriate variant. The experimental advanced scheduler is not a universal arbitrary-model profile.
Why are my Krea 2 RAW dimensions different in schedule_info?
RAW dimensions are normalized to a supported grid. Compare the requested and effective dimensions shown in schedule_info.
Are the model-free workflow examples proof of image quality?
No. They demonstrate schedule construction, not model inference or finished-image quality.
Learn More about ComfyUI-Sigmax
- Compatibility documents host versions and model-specific support boundaries. Consult it when updating ComfyUI or using MiniMax H3 native schedulers.
- Original Qwen Image native example, Qwen Image 2.1 native example, and Qwen Image 2.1 dynamic example show how those schedules are constructed without loading a model.
- ComfyUI’s H3 model registration and native H3 sampling node provide technical detail if you need to understand how the qualified host handles H3 shifts.
- The Qwen Research License explains the separate usage terms for Qwen Image 2.1 model use.
