PixelUMM

PixelUMM: Encoder-Free Unified Image and Video Understanding and Generation

Cong Wei ^1,2   Xuanchi Ren ¹   Bryan Chu ¹   Weiming Ren ²
Huan Ling ¹   Jiahui Huang ¹   Laura Leal-Taixé ¹   Sanja Fidler ¹
Wenhu Chen ²   Zian Wang ¹   Jay Zhangjie Wu ¹

¹ NVIDIA ² University of Waterloo

PixelUMM reads and writes raw pixels with a single decoder-only Transformer — no VAE and no vision encoder. Images become 16×16 patches and videos become 4-frame tubes; a Qwen3-8B backbone with separate understanding and generation experts shares one self-attention across text, clean pixels, and noisy pixels.

Quick start

This repository provides inference code for image and video generation and understanding, plus a four-step toy training example. S8-F22-R05 is the default checkpoint after Stage 2 training and is the checkpoint used in the paper evaluation. S8-F18-R01 is an intermediate checkpoint that received 10K additional fine-tuning steps at 480p and 720p short-side settings. It generally gives slightly better text-to-video results than the default checkpoint. Model weights and toy data are downloaded separately from this source repository.

1. Clone the repository and install the CUDA environment using ENVIRONMENT.md. For T2V, also set up the separate Cosmos guardrail environment; checks are on by default.
2. Download S8-F22-R05 from nvidia/PixelUMM and the Qwen3-8B config/tokenizer files using CHECKPOINT.md. The PixelUMM checkpoint contains the learned language-model weights; separate Qwen weight shards are not needed.
3. Set the local paths and check the checkpoint before running on a GPU:

export PIXELUMM_CKPT=/absolute/path/to/PixelUMM/S8-F22-R05
export PIXELUMM_QWEN_DIR=/absolute/path/to/qwen3-8b-config-tokenizer
export PIXELUMM_OUTPUT=/absolute/path/to/pixelumm-results

CUDA_VISIBLE_DEVICES="" python check_checkpoint.py \
  --checkpoint "$PIXELUMM_CKPT" --llm-path "$PIXELUMM_QWEN_DIR"

check_checkpoint.py checks checkpoint completeness and tensor compatibility without loading the full model on a GPU. Run the examples below on a GPU allocated to you; set CUDA_VISIBLE_DEVICES accordingly.

Inference

For the default T2V safety checks, first request access to nvidia/Cosmos-1.0-Guardrail, accept its access conditions, and sign in with the approved Hugging Face account as described in GUARDRAILS.md. A Hugging Face login alone does not grant access to this separately gated model. Its weights are downloaded on first use.

# Text to image
python inference.py \
  --checkpoint "$PIXELUMM_CKPT" --llm-path "$PIXELUMM_QWEN_DIR" \
  --task t2i --prompt "A red panda reading beside a window" \
  --height 256 --width 256 --seed 4396 \
  --output "$PIXELUMM_OUTPUT/t2i.png"

# Text to video: 96 frames, 24 FPS
python inference.py \
  --checkpoint "$PIXELUMM_CKPT" --llm-path "$PIXELUMM_QWEN_DIR" \
  --task t2v \
  --prompt "A golden retriever runs across a grassy field in bright daylight. The camera tracks smoothly beside the dog." \
  --height 176 --width 320 --frames 96 --fps 24 --steps 35 \
  --negative-prompt-file experiments/s8_f22_r07/t2v_negative_prompt.txt \
  --output "$PIXELUMM_OUTPUT/t2v.mp4"

# Image understanding
python inference.py \
  --checkpoint "$PIXELUMM_CKPT" --llm-path "$PIXELUMM_QWEN_DIR" \
  --task image-vlm --image /absolute/path/to/photo.jpg \
  --prompt "Describe the image." \
  --output "$PIXELUMM_OUTPUT/image-answer.txt"

# Video understanding
python inference.py \
  --checkpoint "$PIXELUMM_CKPT" --llm-path "$PIXELUMM_QWEN_DIR" \
  --task video-vlm --video /absolute/path/to/clip.mp4 \
  --prompt "Describe the video." \
  --output "$PIXELUMM_OUTPUT/video-answer.txt"

The default F22 checkpoint supports all four tasks. S8-F18-R01 supports image/video generation and image understanding, but not video understanding. Its separate download, model profile, and 832×464 T2V inference command are in CHECKPOINT.md.

For a YAML prompt batch, use inference_batch.py --help. The output directory must be new for each run. This batch entrypoint does not run guardrails; the default checks currently apply to single-command T2V only. For benchmark evaluation, see EVAL.md. Pass --no-guardrails to the single-command T2V entrypoint only when you explicitly intend to skip its safety checks.

Four-step toy training

Download and verify the toy data package, then run the complete four-task example in TRAIN.md. The example starts from the F22 checkpoint and saves model weights that inference.py can load. It is a functional training example, not a full training recipe or an optimizer-state resume. Plan for seven GPUs with at least 48 GiB each and sufficient disk space for an approximately 61 GB output checkpoint.

Repository layout

• modeling/: model architecture and generation modules.
• data/: local multimodal dataset and preprocessing.
• train/, train_toy.py: model loading and toy training.
• inference.py, inference_batch.py: inference entrypoints.
• experiments/: model profiles and generation settings.
• eval/: optional evaluation tools.

Most source files are licensed under Apache-2.0. File-specific notices are retained where upstream code has different terms; in particular, modeling/pixelumm/modeling_utils.py retains its DiT-derived CC BY-NC 4.0 notice. Model weights have separate terms in the model repository.

Citation

If you use PixelUMM, please cite our paper:

@misc{wei2026pixelummencoderfreeunifiedimage,
  title={PixelUMM: Encoder-Free Unified Image and Video Understanding and Generation},
  ={Cong Wei and Xuanchi Ren and Bryan Chu and Weiming Ren and Huan Ling and Jiahui Huang and Laura Leal-Taixé and Sanja Fidler and Wenhu Chen and Zian Wang and Jay Zhangjie Wu},
  year={2026},
  eprint={2609.38597},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2609.38597},
}