ComfyUI-Sigmax

ComfyUI-Sigmax provides model-aware sigma schedules for ComfyUI, with supported Krea 2, Z-Image, FLUX.1-schnell, Qwen Image and Qwen Image 2.1, SD3, AuraFlow v0.2, Lumina-Image 2.0, HunyuanImage 2.1, MiniMax H3 Base FL2VA/Ref2VA plus experimental community Turbo recipes, Anima, Wan 2.1/2.2, Wan Animate 2, and LTX profiles plus editing tools.

Last Update

Version 1.1.3 — recent features and improvements

Qwen Image schedules

• Added a dedicated Qwen Image 2.1 node with native and dynamic schedules, explicit target dimensions or tokens, and model-free examples.
• Added an explicit native mode for original Qwen Image while preserving existing saved modes, defaults, and numerical behavior.

Video model profiles

• Expanded Wan profiles with FLF2V, VACE, S2V, Animate, and Wan Animate 2 Base/Distilled modes.
• Added the ComfyUI-optimized Wan Animate 2 recipe for the matching model and LoRA: 480P, LCM, Simple, 6 steps, CFG 1.0, and shift 5.0.
• Added experimental MiniMax H3 Turbo recipes and a ten-choice scheduler menu with explicit model-shift ownership.

Host compatibility and reliability

• Requalified MiniMax H3 native schedulers on ComfyUI 0.39.0 while retaining pinned-host support and rejecting mismatched model shifts.
• Added independent Qwen numerical checks and current/pinned-host schedule validation.
• Improved cross-platform host-log path redaction and release-artifact consistency checks.

Sampler research contracts

• Expanded deterministic and seeded stochastic Flow Euler checks, sampler-state receipts, and advanced-workflow compatibility decisions.
• These remain internal research contracts; no replacement public sampler is registered.

Table of contents

• Last Update
• Features
• Installation: ComfyUI Manager · Git
• Use in ComfyUI
  • Build a model schedule
    • Image model families
    • Video and audio-video model families
  • Inspect or modify a schedule
• Update or remove
• Troubleshooting

Features

• Explicit model and variant selection with no silent generic fallback.
• Supported Krea 2, Z-Image, FLUX.1-schnell, Qwen Image and Qwen Image 2.1, SD3, AuraFlow v0.2, Lumina-Image 2.0, HunyuanImage 2.1, MiniMax H3 Base FL2VA/Ref2VA, Anima, Wan 2.1/2.2, Wan Animate 2, and LTX recipes.
• Twenty-five namespaced nodes for schedule construction, model-aware selection, inspection, comparison, editing, checkpoint evidence, and experimental Krea 2 conditioning.
• Schedule slicing, concatenation, resampling, inspection, and comparison.
• Experimental Krea 2 CONDITIONING tap rebalancing for explicitly selected RAW or Turbo workflows, with fixed RMS preservation and no scheduler/model patching.
• MiniMax H3 Base workflow construction with explicit FL2VA/Ref2VA selection, upstream model-shift integration, and model-free /object_info schema preflight.
• Experimental MiniMax H3 Turbo recipe selection for source-qualified community lanes. The selector constructs recipe-owned sigmas and receipts only; it does not load or patch a LoRA, attention backend, or model.
• A BasicScheduler-style MiniMax H3 scheduler menu with one dependency-free h3_endpoint default and nine experimental ComfyUI-native choices that require an already-shifted H3 MODEL.
• Internal, dependency-free sampler-state and Flow Euler contracts for deterministic/stochastic research; they are not registered public sampler nodes and do not replace ComfyUI's sampler.
• Checkpoint header inspection without loading model weights.
• Versioned schedule information and fingerprints for reproducible workflows.
• No mandatory third-party runtime dependencies beyond the libraries provided by ComfyUI.

Sigmax builds SIGMAS for ComfyUI custom sampling. It does not replace the sampler and does not download models.

Installation

ComfyUI Manager

Search for ComfyUI-Sigmax in ComfyUI Manager, install it, and restart ComfyUI. If it is not available in Manager, use the Git installation below.

Git

Stop ComfyUI, open a terminal in ComfyUI/custom_nodes, and run:

git clone https://github.com/rookiestar28/ComfyUI-Sigmax comfyui-sigmax

Restart ComfyUI. The expected package entry point is:

ComfyUI/custom_nodes/comfyui-sigmax/__init__.py

Python 3.10 or newer and ComfyUI 0.29.0 or newer are required. Model-specific host requirements and validated versions are listed in Compatibility. Do not install the repository's development or parity dependencies into ComfyUI just to use the nodes.

Use in ComfyUI

Search the node menu for Sigmax. The package registers 25 namespaced nodes.

Build a model schedule

Image model families

Model	Node	Recommended settings
Krea 2 Turbo	Sigmax.Krea2SigmaScheduler	Turbo, 8 steps, Euler, CFG 1.0
Krea 2 RAW	Sigmax.Krea2SigmaScheduler	RAW, 52 steps with CFG 4.5, or 28 steps with CFG 5.5
Krea 2 RAW + RAW-to-Turbo LoRA	Sigmax.Krea2SigmaScheduler	Experimental RAW/Turbo mu; user-selected steps; community starting point: 12 steps, Euler, CFG 1.0
Z-Image Base	Sigmax.ZImageSigmaScheduler	Base, 28-50 steps, default 50, CFG 4.0
Z-Image Turbo	Sigmax.ZImageSigmaScheduler	Turbo, 8 steps, CFG 1.0
FLUX.1-schnell	Sigmax.Flux1SchnellSigmaScheduler	1-4 steps, default 4, CFG 1.0
Original Qwen Image	Sigmax.QwenImageSigmaScheduler	Explicit Comfy Native uses exponential mu 1.15 and the 10000-entry simple table; legacy Comfy Fixed and Diffusers Dynamic remain available; 50-step compatibility default, host true CFG 4.0
Qwen Image 2.1	Sigmax.QwenImage21SigmaScheduler	Comfy Native: 25 steps, mu 0.69, Euler/Simple, true CFG 1.0. Diffusers Dynamic: 40 steps, explicit target dimensions or tokens, true CFG 1.0
Original Stable Diffusion 3	Sigmax.SD3SigmaScheduler	Select Publisher Reference (1.0) at 50 steps or Comfy/Diffusers Fixed (3.0) at 28 steps; source mode is required
Original AuraFlow v0.2	Sigmax.AuraFlowSigmaScheduler	Official Fixed (1.73), 50 steps, CFG 3.5; source mode is explicit
Lumina-Image 2.0	Sigmax.Lumina2SigmaScheduler	Official Fixed (6.0), 50 steps, CFG 4.0; source mode is explicit
HunyuanImage 2.1	Sigmax.HunyuanImage21SigmaScheduler	Base (5.0), 50 steps, CFG 3.5, or Distilled (4.0), 8 steps, CFG 3.25; variant is explicit
Anima Base v1.0 / Aesthetic / Turbo	Sigmax.AnimaSigmaScheduler	Base, 30-50 steps, default 50, CFG 4.5; Aesthetic uses the same recipe; Turbo, 8-12 steps, CFG 1.0; fixed shift 3.0

Connect an image scheduler's SIGMAS directly to a custom-sampling path that accepts external sigmas. Keep strict_official enabled for official variants and use schedule_info to confirm the recipe, dimensions, step range, and warnings. Do not pass the result through another scheduler or apply another time shift.

• Krea 2: Select Turbo or RAW explicitly and enter the actual output width/height. Sigmax.ModelAwareSigmaScheduler may identify only the family, so resolve an ambiguous Auto result manually. Sigmax.Krea2ConditioningRebalance is an optional experimental CONDITIONING transform with identity at strength 0; select RAW/Turbo and its profile explicitly. The RAW-to-Turbo LoRA schedule modes force strict_official off, do not load or scale a LoRA, and are only for a compatible LoRA applied to RAW.

• Z-Image: Select Base or Turbo explicitly; the node supplies the validated schedule only and does not load a checkpoint or choose a sampler.

• FLUX.1-schnell: Use the dedicated 1-4-step schedule for the schnell family; other FLUX variants are not implied.

• Qwen Image: Comfy Native selects the source-qualified exponential mu 1.15 schedule. The saved/default Comfy Fixed choice preserves the legacy direct-ratio 1.15 formula; it differs numerically from native ComfyUI. Diffusers Dynamic requires explicit packed image_seq_len. Select migration explicitly and use exactly one external-sigma source, without applying another shift or model-sampling patch. The packaged native model-free example constructs sigmas only. Image-quality parity is outside scope.

• Qwen Image 2.1: Use its separate node and profiles. Native mode requires None (Native) with width, height and tokens all zero. Dynamic mode requires Dimensions with positive width/height divisible by 32 and tokens zero, or Target Tokens with positive image_seq_len and dimensions zero. Target tokens are (width / 16) * (height / 16), without packed patches or reference-image tokens. Dynamic shift is unclamped and stretches the final nonzero sigma to 0.02 before appending zero. Keep exactly one external-sigma source. The native and dynamic examples construct schedules only. Current ComfyUI 0.39 model-free execution and numerical validation do not establish model inference or image quality. Model use is governed separately by the publisher's Qwen Research License, which restricts use to noncommercial research/evaluation and requires separate authorization for commercial use.

  Changing the source mode does not change steps automatically. With strict_official enabled, use 25 steps for Comfy Native or 40 for Diffusers Dynamic; disable it explicitly for a modified step count. Leave inactive geometry/token controls at zero.

• Stable Diffusion 3: Only original SD3 Medium is covered. Choose the publisher 1.0 or pinned ComfyUI/Diffusers 3.0 source mode explicitly; SD3.5, Turbo, ControlNet, execution, and quality claims are excluded.

• AuraFlow: Support is limited to original fal AuraFlow v0.2 with fixed ratio 1.73 and 50 steps; other versions, finetunes, dynamic shifts, execution, and quality claims are excluded.

• Lumina-Image 2.0: Support is limited to the original Alpha-VLLM text-to-image schedule with fixed ratio 6.0 and 50 steps; video, mGPT, editing paths, dynamic shifts, execution, and quality claims are excluded.

• HunyuanImage 2.1: Base and Distilled use explicit direct-ratio paths (5.0/4.0). Base is the pinned ComfyUI-compatible lane; Distilled remains publisher-schedule-only. Refiner, encoders, conditioning, weights, and quality claims are excluded.

• Anima: Select Base, Aesthetic, or Turbo explicitly. Aesthetic shares the Base recipe; all three paths construct schedules only and do not load model components.

Video and audio-video model families

Model	Node	Recommended settings
MiniMax H3 Base / experimental Turbo	Sigmax.MiniMaxH3SigmaScheduler	Select FL2VA or Ref2VA explicitly. h3_endpoint remains the default pure schedule; nine additional ComfyUI-native scheduler choices are experimental and require the already-shifted H3 MODEL. The optional Turbo selector exposes source-qualified community recipes and is readiness-only.
Wan 2.1 T2V / I2V	Sigmax.WanSigmaScheduler	Select generation, task, source, and resolution explicitly; official T2V 50 steps (5.0), official I2V 480P/720P 40 steps (3.0/5.0)
Wan 2.1 FLF2V / VACE	Sigmax.WanSigmaScheduler	Official-native FLF2V 14B 720P uses 50 steps (16.0); VACE 1.3B/14B use 50 steps (16.0)
Wan 2.2 TI2V 5B / A14B T2V/I2V	Sigmax.WanSigmaScheduler	Native or Diffusers-reference source lanes; TI2V 5B uses 5.0; A14B T2V/I2V use 12.0/5.0 with caller-owned boundary metadata
Wan 2.2 S2V / Animate	Sigmax.WanSigmaScheduler	Official-native S2V 14B uses 40 steps (3.0); Animate 14B uses 20 steps (5.0)
Wan Animate 2 Base / Distilled	Sigmax.WanSigmaScheduler	Official-native Base uses 40 steps (5.0); Distilled uses 10 steps (5.0); select the task explicitly
Wan Animate 2 Comfy optimized	Sigmax.WanSigmaScheduler	Select Animate Optimized, ComfyUI native, and 480P; use the matching model and step-distilled LoRA with LCM, the supplied Simple sigmas, 6 steps, CFG 1.0, and shift 5.0
LTXV 0.9.8 / LTX-2 19B / LTX-2.3 22B	Sigmax.LTXSigmaScheduler	Dev adaptive token shift (20/40/30 default steps) or explicit LTX-2/LTX-2.3 distilled Stage 1/2 vectors; generation and stage are explicit

Connect a video scheduler's SIGMAS directly to the matching custom-sampling path. Do not add another scheduler or time shift, and inspect schedule_info for the selected generation mode, stage, resolution, boundary ownership, and warnings.

• MiniMax H3: Select Base FL2VA or Ref2VA explicitly. The scheduler menu contains h3_endpoint, simple, sgm_uniform, karras, exponential, ddim_uniform, beta, normal, linear_quadratic, and kl_optimal. The default h3_endpoint path is the existing pure endpoint-inclusive schedule and needs no MODEL. Every other choice is an experimental ComfyUI-native lane: connect the already-shifted H3 MODEL, keep its video/audio shifts aligned with the selected Base or Turbo recipe, and do not add a separate BasicScheduler or another time shift. Use the upstream MiniMaxH3SigmaShift node (also displayed as ModelSamplingMiniMaxH3) when the host or recipe needs an explicit shift configuration. These native choices are functional compatibility options, not MiniMax recommendations or quality, speed, memory, NFE, or acceleration claims. Exact 0.30.0, 0.32.0 and 0.39.0 host qualification is described in Compatibility.

  On qualified ComfyUI 0.39.0, native H3 registration supplies the Base defaults video_shift=12.0 and audio_shift=3.0; an explicit ModelSamplingMiniMaxH3 node is normally unnecessary when those values already match the selected recipe. When an explicit configuration is needed, connect that node after the final model branch (after any LoraLoaderModelOnly or model switch) and before the guider or scheduler. The 768p FL2VA 4-step recipe requires 6.0/3.0. The node changes model sampling metadata; it does not load a Turbo LoRA or an attention backend. See the qualified ComfyUI H3 model registration and native H3 sampling node for the host-owned behavior.

  The optional turbo selector exposes four source-qualified community recipes. The selector only constructs recipe-owned sigmas and readiness receipts; it does not load or patch a LoRA, attention backend, or model. Load the matching LoRA separately and keep the workflow's task, resolution, step count, and shifts aligned with this table:

  Recipe	Steps	Video/audio shift	Task / 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

  These Turbo recipes are experimental community compatibility lanes. They make no official-method, quality, speed, memory, NFE, or acceleration claim. For an official current BasicScheduler workflow, the model path is normally UNETLoader -> (optional) LoraLoaderModelOnly -> (optional, only when needed) ModelSamplingMiniMaxH3 -> BasicGuider/BasicScheduler. When using Sigmax's scheduler node, connect its SIGMAS output directly to the custom-sampling path; do not add a second scheduler or shift.

• Wan 2.1/2.2 and Wan Animate 2: Select generation, task, source, and resolution explicitly. Wan 2.1 I2V requires 480P or 720P; Wan 2.2 A14B boundaries are caller-owned metadata and never route experts. FLF2V, VACE, S2V, Animate, and Wan Animate 2 Base/Distilled rows are official-native schedule math only; Diffusers-reference lanes describe scheduler construction only. The separate Comfy optimized lane reproduces the first-party 480P LCM + simple + 6 recipe for its exact model and LoRA. Its 81-frame value is tested chunk guidance, not a hard host-node maximum or automatic chunking feature. Execution, weight loading, expert routing, conditioning, and video-quality parity are excluded.

• LTX: Select LTXV 0.9.8, LTX-2 19B, or LTX-2.3 22B plus generation/stage explicitly. Dev mode derives one token-count shift; distilled modes use immutable publisher vectors. Sigmax does not load video weights, run encoders, or claim video-quality parity.

Inspect or modify a schedule

• Sigmax.ProfileInspector and Sigmax.ScheduleInspector show profile and schedule details.
• Sigmax.ScheduleComparison compares two schedules.
• Sigmax.ScheduleSlice, Sigmax.ScheduleConcatenate, and Sigmax.ScheduleResample create a modified schedule while retaining traceable schedule information.
• Sigmax.CheckpointEvidenceInspector reads only the bounded header of a local .safetensors file. Its suggested variant is advisory, not confirmation.
• Sigmax.AdvancedFlowMatchScheduler is an experimental schedule constructor, not a universal model profile or sampler.

Update or remove

For a Git installation:

cd ComfyUI/custom_nodes/comfyui-sigmax
git pull --ff-only

Restart ComfyUI after updating. Back up important workflows before changing versions. To remove Sigmax, stop ComfyUI, remove or move only the comfyui-sigmax directory, and restart ComfyUI.

Troubleshooting

Problem	Resolution
Nodes do not appear	Restart ComfyUI, inspect the startup log, and confirm there is only one direct comfyui-sigmax folder under custom_nodes.
Import fails	Confirm Python 3.10+ and ComfyUI 0.29.0+, then remove duplicate or nested installations.
Auto rejects Krea 2	Select Turbo or RAW explicitly. Do not rely on the filename alone.
The schedule appears shifted twice	Remove the second scheduler or model time shift; use the Sigmax SIGMAS output once.
Krea 2 RAW dimensions differ	Check the requested and effective dimensions in schedule_info; RAW dimensions are normalized to the supported grid.
Qwen Image 2.1 rejects a mode or target	Set 25 native or 40 dynamic steps with strict_official enabled. Native requires zero dimensions/tokens; dynamic requires exactly one target source with inactive controls zero.
MiniMax H3 native scheduler asks for MODEL or reports a shift mismatch	Connect an already-shifted H3 model. Use 12.0/3.0 for Base or the exact selected Turbo recipe; configure ModelSamplingMiniMaxH3 when needed, without a separate BasicScheduler or second shift.
A newer ComfyUI release behaves differently	Review the supported boundary in Compatibility.

ComfyUI custom nodes execute Python code inside the host process. Install reviewed sources only. Model weights are not included. A bounded local Krea 2 H4 lane verified execution and artifact provenance, but blind scoring was waived; it does not establish image-quality parity or general GPU compatibility. Stochastic sampling, resume behavior, and arbitrary-model compatibility are not claimed.