---
license: other
license_name: openmdw1.1-license
license_link: >-
  https://openmdw.ai/license/1-1/
library_name: cosmos
tags:
  - nvidia
  - cosmos
  - cosmos3
  - image-to-video
  - video-generation
  - fp8
  - quantized
  - modelopt
---

# Cosmos3-Super-Image2Video-4Step — FP8 (community quantization)

This is a community **FP8 (E4M3), weight-only** quantization of NVIDIA's
[`nvidia/Cosmos3-Super-Image2Video-4Step`](https://huggingface.co/nvidia/Cosmos3-Super-Image2Video-4Step) —
a 64B-parameter image-to-video model, DMD2-distilled for 4-step, CFG-free
generation — produced with [NVIDIA TensorRT Model Optimizer](https://github.com/NVIDIA/TensorRT-Model-Optimizer)
(`nvidia-modelopt`).

**Model Developer (original model): NVIDIA.** This is not an official NVIDIA
release. Everything about the model itself — architecture, training data,
benchmarks, full input/output specs, limitations, and ethical considerations —
is unchanged from NVIDIA's original and is documented on
[their model card](https://huggingface.co/nvidia/Cosmos3-Super-Image2Video-4Step);
it isn't reproduced here. This card only covers what we changed (the
quantization) and how to run *this* checkpoint.

## License

Released under the [OpenMDW-1.1](https://openmdw.ai/license/1-1/) license, same
as the original model — see [`LICENSE`](LICENSE) in this repo.

## Quantization Notes

**Recipe:** per-tensor E4M3 weights on the 64 transformer blocks' attention/MLP
linears. Activations, embeddings, norms, the bundled Qwen3 reasoner head,
`time_embedder`, and `proj_in`/`proj_out` are left in BF16. Weight-only, so no
calibration forward pass was needed. Per NVIDIA's own note, FP8 is not
officially tested/supported for this model — treat this as best-effort.

**Loading gotchas — read before using with `diffusers`:** as of
`diffusers==0.39.0.dev0`, plain `Cosmos3OmniPipeline` usage has **two** bugs
against this checkpoint. Both are fixed automatically by the serving script
below (`serve_cosmos3_i2v4step_diffusers.py`); if you're writing your own
loading code, patch the scheduler yourself as shown.

1. **Wrong step count / CFG on.** `Cosmos3OmniPipeline.__call__` does not read
   this checkpoint's `scheduler/scheduler_config.json` →
   `fixed_step_sampler_config.t_list` (the trained 4-step sde schedule). It
   silently falls back to the pipeline's generic defaults
   (`num_inference_steps=35`, `guidance_scale=6.0`), which is off-distribution
   for a DMD2-distilled 4-step checkpoint. (`vLLM-Omni` parses this file
   correctly; this gap is specific to the plain `diffusers` PyTorch path.)

2. **Image conditioning gets destroyed into colorful static (critical).** This
   checkpoint's scheduler ships `stochastic_sampling=True` (SDE sampling).
   Cosmos3 anchors the conditioning frame by zeroing the model's predicted
   velocity there — under a *deterministic* step that means "leave this frame
   unchanged," but `FlowMatchEulerDiscreteScheduler`'s SDE branch computes
   `x0 = sample - current_sigma * model_output` (= `sample`, since velocity is
   0) then `prev_sample = (1 - next_sigma) * x0 + next_sigma * randn_tensor(...)`
   — it re-noises by `next_sigma` **regardless of velocity**. Across this
   checkpoint's 4 steps that compounds to ~99.6% fresh noise in the
   conditioned frame: your input image comes out as colorful static while the
   genuinely-denoised motion frames still look like plausible video (this
   combination — garbage first frame, coherent-but-drifted rest — is exactly
   how it presents). Disabling `stochastic_sampling` restores the correct
   zero-velocity-is-a-no-op behavior. Confirmed by direct A/B render, same
   seed/image/prompt, only this flag changed.

```python
def force_fixed_step_schedule(scheduler):
    t_list = scheduler.config.fixed_step_sampler_config["t_list"]
    orig = scheduler.set_timesteps
    scheduler.set_timesteps = lambda num_inference_steps=None, device=None, **_: (
        orig(sigmas=list(t_list), device=device)
    )
    if scheduler.config.stochastic_sampling:
        scheduler.register_to_config(stochastic_sampling=False)

pipe = ...  # Cosmos3OmniPipeline.from_pretrained(...)
force_fixed_step_schedule(pipe.scheduler)

result = pipe(prompt=..., image=..., guidance_scale=1.0, num_inference_steps=4, ...)
# guidance_scale=1.0 disables CFG; num_inference_steps is a no-op once patched
```

Verified against this exact repo: the patch produces a 4-iteration denoising
loop (not 35) with CFG off and a correctly-preserved conditioning frame
(not colorful static), matching the checkpoint's trained regime. The
save/restore round trip (`transformer/modelopt_state.pth`) reloads correctly
with 896 quantized weight wrappers active.

## Usage: Run Inference (single GPU, `diffusers`)

This repo is a `diffusers`-loadable repackage (`transformer/modelopt_state.pth`),
not NVIDIA's `vLLM-Omni` deployment export format. If you have a vLLM-Omni
cluster, use NVIDIA's original BF16 checkpoint and card instead. For everyone
else — anyone running this FP8 checkpoint on a single GPU — use the FastAPI
server included in this repo, which loads the checkpoint, applies the 4-step
scheduler fix above automatically, and exposes a plain HTTP endpoint.

### 1. Install

Requires a `diffusers` build with Cosmos3 support — not yet in a PyPI release
as of this writing, so install from the exact commit this checkpoint was
produced and validated against:

```bash
pip install "git+https://github.com/huggingface/diffusers.git@2c7efb95349296cf6bcce981ea036275a82a94df"
pip install nvidia-modelopt accelerate torch fastapi uvicorn python-multipart
```

### 2. Download this repo (weights + scripts together)

```bash
hf download prometheusAIR/Cosmos3-Super-Image2Video-4Step-FP8 \
    --local-dir Cosmos3-Super-Image2Video-4Step-FP8
cd Cosmos3-Super-Image2Video-4Step-FP8
```

This pulls the whole ~66GB repo — FP8 weights, VAE, tokenizer, and the `.py`
scripts side by side, the same download either way.

### 3. Serve it

```bash
CUDA_VISIBLE_DEVICES=0 python serve_cosmos3_i2v4step_diffusers.py --repo .
```

One GPU with roughly 70GB+ free memory (RTX PRO 6000 96GB, H100/H200,
A100-80GB, etc.) is enough — no multi-GPU sharding required for this FP8
checkpoint. The server listens on `http://localhost:8000` once it's done
loading.

### 4. Use it

```bash
curl -s -X POST http://localhost:8000/animate \
    -F image=@your_first_frame.png \
    -F 'prompt=The robotic arm slowly lowers its gripper toward the objects and holds. Static camera.' \
    -F num_frames=49 -F fps=24 \
    --output clip.mp4
```

`num_frames` / `fps` / `height` / `width` are adjustable. `num_inference_steps`
and `guidance_scale` are intentionally not exposed — this checkpoint's 4-step,
CFG-free schedule is fixed and applied automatically by the server.

### Other scripts in this repo

- `quantize_cosmos3_i2v4step_streaming.py` — reproduces this FP8 quantization
  from NVIDIA's BF16 source checkpoint.
- `repackage_for_hf_i2v4step.py` — rebuilds this diffusers-loadable repo format
  from a quantized transformer.
- `load_cosmos3_modelopt.py` — the underlying loader `serve_cosmos3_i2v4step_diffusers.py`
  uses; import `load_pipe(...)` directly if you want a `pipe` object instead of
  an HTTP server.
- `validate_cosmos3_i2v4step_fp8.py` — a minimal standalone image→video smoke
  test against this checkpoint.

## Responsible Use

See NVIDIA's [Bias](BIAS.md), [Explainability](EXPLAINABILITY.md),
[Safety & Security](SAFETY.md), and [Privacy](PRIVACY.md) subcards (included
in this repo), and the Limitations / Ethical Considerations sections of
[NVIDIA's original model card](https://huggingface.co/nvidia/Cosmos3-Super-Image2Video-4Step).
Report security vulnerabilities [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).