speed-generator

name

speed-generator

description

Generate images and videos with SPEED (Spectral Progressive Diffusion), a training-free method that runs diffusion at progressively increasing resolution for a large speedup over full-resolution sampling. Use this skill whenever the user wants to generate an image with FLUX.1-dev, a pixel-space image with PixelGen, or a video with WAN 2.1 — and especially when they mention SPEED, progressive resolution, faster/efficient/accelerated diffusion sampling, spectral noise expansion, timestep alignment, DCT/DWT/FFT resolution transitions, or running latent_image_gen / pixel_image_gen / latent_video_gen. Reach for it for any "generate an image/video efficiently" request here even if the user never says the word "SPEED".

SPEED Generator

Spectral Progressive Diffusion (SPEED) generates an image or video by running the denoising trajectory at increasing spatial resolution instead of full resolution the whole way. At each transition the latent is spectrally expanded (DCT/DWT/FFT) with noise-filled high frequencies and the flow-matching timestep is aligned, so the model continues seamlessly at the larger resolution.

This skill drives three standalone CLI generators bundled in scripts/:

Script	Model	Output	Space
latent_image_gen.py	FLUX.1-dev	PNG image	latent image
pixel_image_gen.py	PixelGen	PNG image	pixel-space
latent_video_gen.py	WAN 2.1	MP4 video	latent video

utils.py holds the shared spectral math; configs.yaml holds per-model checkpoint paths, power-spectrum fits, and defaults.

When to use which generator

• FLUX.1-dev (latent_image_gen.py) — the default for a normal text-to-image request. Fast, high quality, guidance-distilled.
• PixelGen (pixel_image_gen.py) — pixel-space generation (square only). Use only when the user specifically wants PixelGen.
• WAN 2.1 (latent_video_gen.py) — text-to-video.

If the user just says "generate an image of X efficiently" with no model named, use FLUX.1-dev.

Setup (do this once per session before generating)

1. Environment: The scripts do from utils import ..., so the bundled scripts/ directory must be on PYTHONPATH. Use the correct environment with all dependencies from the speed/ repo installed.

2. Checkpoints via env vars (configs.yaml expands ${...} placeholders; an unset one raises a clear KeyError). Set only what the chosen model needs:

   export FLUX_DIR=/local/howard/efficient_diffusion/wavelet_diffusion/models/FLUX.1-dev
   export PIXELGEN_REPO=/home/howard/efficient_diffusion/PixelGen
   export PIXELGEN_CKPT=/local/howard/efficient_diffusion/models/pixelgen/PixelGen_XXL_T2I.ckpt
   export PIXELGEN_CONFIG=$PIXELGEN_REPO/configs_t2i/sft_res512.yaml
   export WAN_PATH=/home/howard/efficient_diffusion/Wan2.1
   export WAN_CKPT=/local/howard/efficient_diffusion/models/Wan2.1-T2V-1.3B
   

Running a generation

Always pass PYTHONPATH=<skill>/scripts and CUDA_VISIBLE_DEVICES=$GPU. Use --device cuda:0 (it maps to the pinned GPU) and --progress for a denoising progress bar.

FLUX image (the common case):

CUDA_VISIBLE_DEVICES=$GPU PYTHONPATH=scripts \
python scripts/latent_image_gen.py \
    --prompts "a translucent jellyfish glowing in deep water" \
    --transform dct --scales 0.5 1.0 --delta 0.01 \
    --save_dir ./out_flux --device cuda:0 --progress

PixelGen image:

CUDA_VISIBLE_DEVICES=$GPU PYTHONPATH=scripts \
python scripts/pixel_image_gen.py \
    --prompts "a cute puppy" --transform dct --scales 0.5 1.0 \
    --save_dir ./out_pixelgen --device cuda:0 --progress

WAN video (480p → 720p):

CUDA_VISIBLE_DEVICES=$GPU PYTHONPATH=scripts \
python scripts/latent_video_gen.py \
    --prompts "a dog running in a meadow" \
    --transform dct --scales 480 720 \
    --save_dir ./out_wan --device cuda:0 --progress

Each script prints the output path(s) it wrote (one pNNNN.png / .mp4 per prompt).

Shared flags

Flag	Meaning
--prompts "a" "b" / --prompt_txt f.txt / --prompt_csv f.csv	Prompt input (mutually exclusive; CSV column prompt).
--transform {dct,dwt,fft}	Spectral basis at each transition. Default dct.
--scales ...	Increasing stage sizes ending at full resolution. See below.
--delta	Noise-dominated tolerance for transition scheduling. Default 0.01. Smaller = transition later.
--n_steps, --guidance	Override the per-model defaults in configs.yaml.
--height, --width	Output size (defaults per model: FLUX 1024², PixelGen 512², WAN 720×1280).
--seed, --save_dir, --device, --verbose, --progress	Reproducibility, I/O, logging, progress bar.
WAN extra: --num_frames, --shift · PixelGen extra: --timeshift, --neg_prompt	Model-specific.

--scales — three accepted forms (freely mixed)

The stage list must be strictly increasing and end at full resolution. Each value may be:

• a decimal scale — 0.5 1.0, 0.37 1.0
• a fraction — 1/2 1, 2/3 1, 1/3 2/3 1
• a pixel height — 480 720 (the last value must equal --height)

--scales 1.0 runs the plain full-resolution baseline (no progressive stages).

Transform vs. scales: DCT and FFT accept any ratio between consecutive scales (so 0.37 1.0 is fine). DWT is restricted to a 2× ratio between every consecutive pair (e.g. 0.25 0.5 1.0) because Haar IDWT is structurally a 2× upsample — it raises a clear error otherwise. When in doubt, use dct.

Choosing settings

• Default schedule: --transform dct --scales 0.5 1.0 --delta 0.01. This is the canonical two-stage config and a good starting point.
• More speedup: add an earlier stage, e.g. --scales 0.25 0.5 1.0 (DWT/DCT) or --scales 0.33 0.67 1.0 (DCT/FFT). Lower scales = cheaper early steps but more aggressive; very low first scales can degrade quality.
• Fewer steps = faster but lower quality. --n_steps overrides the model default (FLUX/WAN 50, PixelGen 25).