A ComfyUI implementation of SPEED from Spectral Progressive Diffusion for Efficient Image and Video Generation by Howard Xiao, Brian Chao, Lior Yariv and Gordon Wetzstein (Stanford, 2026). The method is theirs; this pack brings it to ComfyUI's samplers and adds calibration for each model.

"Why make big noise when little noise do trick?"

Make video and image generation faster without retraining.

• The paper's idea: start denoising on a cheaper low-resolution grid, then increase the resolution as finer detail becomes useful.
• This avoids paying for full-resolution compute during the noisiest early steps.
• Runs on: MiniMax-H3 video and flow image models (MOSS Image, Krea 2).
• H3 audio always stays at full resolution.

cd ComfyUI/custom_nodes/
git clone https://github.com/StanLukuvka/ComfyUI-SPEED.git
# restart ComfyUI

1. Swap in the SPEED node.
   • Image models (MOSS Image, Krea 2): swap your KSampler for SPEED — KSampler, using the same wiring as KSampler. Example: image_moss_SPEED.
   • H3: replace SamplerCustomAdvanced with SPEED — Sampler Advanced, using the same wiring. Example: video_h3_SPEED.
2. Fill in noise_amplitude and noise_decay_exponent. They describe how your model, sampler and schedule behave; left at 0, the node stops with an error.
   • Run a Sigma Harvest node once with the same sampler, schedule and size, and copy its numbers (see Sigma Harvest).
3. Set stages and Tolerance (Delta) (Δ below), then queue: Δ 0.005 for video, 0.05 for images.

Same seed and prompt for every run, so native and SPEED can be compared directly. All results use Euler.

Native	SPEED, 4 stages, Δ 0.05 (1.58×)

• Clip: a 5 s 0.5 MP latte-art pour in a dark café, Modal L40S.
• Calibration: one Harvest run (noise_amplitude A 35.4213, noise_decay_exponent β 1.0126).
• Native: 276 s in the Δ 0.005 session and 288 s in the Δ 0.01/0.05 session; each speedup is against its own session.

• None of the ten clips has a colour flash; every run keeps the pour, the latte art and readable lettering.
• All ten GIFs and the original MP4s: evidence/h3_evidence.

MOSS Image, native	MOSS Image, SPEED Automatic (1.26×)

• Scene: a ramen stall at night as a still, RTX 5070 Laptop; speedups use the unrounded times.
• Krea 2 rows are single runs and vary by tens of seconds (13 GB model in 8 GB VRAM).

• Every SPEED run was clean: no seams, banding or colour shifts, and the sign stays legible.
• All runs, including Krea 2 and every stage count: evidence/image_evidence.

The pack has six nodes in two families:

• SPEED — KSampler Sigma Harvest has no denoise: it always calibrates on the full schedule.

Choose how many resolution stages you want:

• Tolerance (Delta) (Δ), noise_amplitude and noise_decay_exponent decide at which step each resolution switch happens.
• A smaller Δ switches later: safer, but it saves less time.
• noise_amplitude and noise_decay_exponent have no defaults; see Sigma Harvest.

For experimenting: pick the switch points yourself with up to four (goal, resolution) pairs.

• For every stage except the last active one, goal is where it ends (a step, or a 0–1 fraction with ratio_mode = ratio).
• The last active stage runs to the end of the schedule.
• Active resolutions must increase, and the last active resolution must be 1.0.
• Set either value to 0 to skip a stage.

• Run Harvest with the same sampler, schedule and step count you will use; its Δ only goes into the paste line.
• Connect calibration to Preview Any and copy the report's three values into Automatic.
• Harvest takes as long as one normal full-resolution render; it does not run the SPEED chain.
• Re-run it when you materially change the checkpoint, sampler, LoRA/addons, scheduler, or step count.

The report also gives r² and health, which say how well the measured curve fits:

For base H3 with Euler, the Quick start reference values skip Harvest. See Evidence for examples of how these settings affect generation.

• Each SPEED stage is its own sampler call, so multistep samplers start their history fresh at every resolution change.
• The shipped calibration and all evidence are Euler; re-run Harvest with the sampler you use.

Most samplers work. The ones that don't, on H3:

• dpm_fast, dpm_adaptive: never finish denoising on H3, with or without SPEED.
• *_cfg_pp except the Euler ones: need a negative prompt, so use CFGGuider, not BasicGuider.
• uni_pc, uni_pc_bh2: break when a stage gets only one step. Use 2 stages, or more steps.
• SDE samplers work but give a different shot from Euler at the same seed.

• Each workflow has Harvest, Automatic and Manual on one set of loaders; only Automatic is active, set to that model's harvested values.

• The image workflows' calibrations were harvested for their own prompt, seed and settings. For your model and settings, run SPEED — KSampler Sigma Harvest and paste its values in.
• At Δ 0.005 the harvested boundary lands in the first few steps of an image schedule, so it saves little; use Δ 0.05, or Manual with the 0.5× stage ending near half the steps.

• Use denoise as in KSampler.
• The first stage starts from the encoded image shrunk to that stage's size.
• Automatic skips coarse stages the shortened schedule has already passed, so a low denoise saves less, while Manual keeps your stages.
• Run Harvest on an empty latent: an encoded image changes what it measures.

• The KSampler nodes need a flow (CONST) model, in Harvest too. Other models, such as EPS or v-prediction checkpoints, are rejected with an error.
• noise_mask / masked denoising is not supported.
• On the Advanced nodes the latent_image must start empty (both the video and audio streams). SPEED's I2V support comes through MiniMax-H3 keyframe conditioning, which is resized and restored separately.
• The Advanced nodes expect the MiniMax-H3 nested latent layout: one video stream plus one audio stream.

If you use SPEED in research, cite the original paper:

@article{xiao2026speed,
  title   = {Spectral Progressive Diffusion for Efficient Image and Video Generation},
  author  = {Xiao, Howard and Chao, Brian and Yariv, Lior and Wetzstein, Gordon},
  journal = {arXiv preprint arXiv:2605.18736},
  year    = {2026}
}

See CHANGELOG.md for the detailed API-level changes from the point where I started tracking them.

PolyForm Noncommercial 1.0.0 — see LICENSE.md. Noncommercial use only.