Disco Diffusion's original checkpoints and CLIP guidance on current PyTorch and NVIDIA GPUs.

---

Disco Diffusion produced dense, fragmented images by making an unconditional ImageNet diffusion model fight a text signal assembled from many random CLIP crops. The models still work, but the notebook depended on a 2021 Python and PyTorch stack and its runtime assumptions do not hold on recent CUDA hardware.

neodisco loads those original checkpoints and the settings JSON files left by the notebook. It keeps the Disco score gradient, cutout schedules, secondary model and init image path while putting precision, randomness and failures behind explicit controls.

• Original checkpoints. It supports OpenAI's 256 model, Katherine Crowson's 512 finetune and the secondary model without converting their state dictionaries.
• Original settings import. Prompt weights, CLIP checkbox names, schedules, eta, skip steps, seed and init settings pass through one validated configuration layer.
• Bounded modern execution. CLIP scoring stays in fp32, UNet AMP is scoped to model forwards, attention is selected per backend, and the web server retains one CLIP bank.
• Reproducible reference mode. One render owns and restores its RNG state. Strict mode moves stochastic cutout augmentation to CPU to avoid nondeterministic CUDA interpolation backward operations.
• Recorded recovery and failures. A NaN CLIP image gradient skips that guidance step as in original Disco, with a warning and saved step list. Non-finite model predictions, infinite gradients or invalid samples still stop the render.

Original Disco compatibility

The current still-image path uses independent cutout draws per CLIP model at its native resolution, the original ResizeRight kernel, the notebook's timestep schedule and its unconditioned prediction for saved images. CLIP draws for each model are batched without sharing draws across models. Secondary guidance runs in fp32. The imported range_scale is preserved, but its contribution through the blended-image derivative is zero, matching the original notebook's graph. This deliberately preserves a reference quirk.

The web example remains 250 configured steps minus 10 skipped, 240 actual iterations. Power schedules such as cut_ic_pow="[1]*400+[2]*600" are supported. --deterministic means repeatable modern execution, not historical notebook emulation: OpenCLIP stays fp32, UNet precision remains configurable, and each render starts its own seed stream. A later image in an old notebook batch cannot be reconstructed from its base seed alone.

See the source audit and compatibility implementation report for oracle coverage, measured speed and remaining numerical differences.

Install

Python 3.10 or newer is declared. The tested environments are listed in the implementation report. Install PyTorch for your CUDA platform first if its standard wheel is not suitable, then install neodisco:

pip install -e .
pip install -e ".[init,webui]"   # LPIPS init guidance and the web application

For development, uv sync --frozen --extra dev --extra webui enforces the checked-in lockfile. The vendored sampling subset from OpenAI guided-diffusion keeps its MIT licence and notice under neodisco/backends/_guided_diffusion.

Environment	Validation
Python 3.12, torch 2.11 + CUDA 12.8, RTX 5090	real 256/512 checkpoint smoke
Python 3.14, torch 2.14, macOS CPU	offline regression tests
Python 3.10+ and torch 2.4+	declared compatibility, not fully measured

Download the 256 or 512 diffusion checkpoint and, for the default guidance path, secondary_model_imagenet_2.pth into weights/disco. The source URLs and expected filenames are recorded in the implementation report.

Use

Run an original settings file:

neodisco --disco-config examples/cobanov-spaceship.json --out spaceship.png

Or provide weighted prompts directly:

neodisco "an ukiyo-e city::1" "neon::-0.2" \
  --image-size 512 --width 1280 --height 768 --steps 250 --eta 0.8 \
  --precision auto --attention sdpa --cut-batch auto --out city.png

Every successful CLI render writes a PNG and a versioned JSON record beside it. The record contains the resolved seed, precision, cut batch, attention and compile mode, plus an init-image hash when used. Input dimensions must be positive multiples of 64.

clamp_max remains Disco's supported guidance control. 0.05 is the default; zero keeps the raw, unclamped Disco score. The former --strength option is rejected because its step-relative formula was tied to an incorrect hand-written sampler update. It was a no-op in the previous reference-DDIM path, so accepting it would be misleading.

An init image controls composition through --skip-steps. A nonzero --init-scale adds LPIPS(x_in, init).sum() in Disco's [-1, 1] blended-image domain and requires the init extra. Missing LPIPS is an error. Example:

neodisco --disco-config settings.json --init-image small.png \
  --skip-steps 125 --init-scale 1000 --out large.png

Run the existing web interface with the same validation and runtime policy:

neodisco-web --weights weights/disco --out outputs \
  --device auto --precision auto --attention sdpa --cut-batch auto

The UI remains at http://127.0.0.1:7870. The server serializes GPU jobs, reports input errors as HTTP 400 responses, preserves previews and history, and restores both original Disco JSON and versioned result JSON. Start it with --deterministic when the web process must accept strict reference jobs; this prepares CUDA before model loading.

Benchmark

The smoke profile runs the real 256 checkpoint without reducing work silently. The representative profile includes a packaged 1280x768, 250-step config with three CLIP models, four cutout draws and eta 0.8:

neodisco-benchmark --profile smoke --weights weights/disco

neodisco-benchmark --profile representative \
  --config examples/cobanov-spaceship.json --weights weights/disco \
  --variant corrected-eager --variant compiled --compile-mode default \
  --runs 3 --warmup 1

Performance measurements from the earlier shared-cutout implementation do not describe the current compatibility path. The compatibility report records fresh RTX 5090 timings with independent cutouts, ResizeRight and fp32 secondary.

Each run directory contains PNGs and a machine-readable manifest with code provenance, package and GPU versions, checkpoint hashes, effective settings, actual executed steps, load and render timings, allocated and reserved peaks, CLIP distance and paired image metrics. Add --profile-stages for a Chrome trace. A failed render writes a failed manifest and exits nonzero, without publishing a speed figure.

How it is built

CLIP image and text operations explicitly disable outer autocast. bf16 and fp16 are accepted only on compatible CUDA devices; fp16 remains opt-in because wide attention can overflow. torch.compile targets the stable UNet forward and defaults to eager. Compiler failures may fall back to eager and are recorded; OOM and numerical failures propagate.

Default CUDA augmentation can differ between repeated runs even with identical random draws because interpolation backward is nondeterministic. --deterministic preserves the transforms but computes them on CPU, which is slower. Reproducibility is promised within one supported software and hardware configuration, not across PyTorch releases or GPU architectures.

CPU rendering is useful for small tests, but full checkpoint rendering is intended for CUDA. Animation, training, replacement latent models, AMD/MPS production support and distributed inference are outside this release.

Licence

MIT. Vendored guided-diffusion code is MIT, copyright OpenAI. Vendored ResizeRight is MIT, copyright Assaf Shocher; its pinned source and licence are included in the package.