primecai/fixer-concat-self-forcing

Fixer Concat — Self-Forcing student

Concat student at SF iteration 1,400, after 1,000 causal flow-matching warm-start updates. This is an intermediate checkpoint from the Fixer pilot, exported from the latest complete checkpoint available when this upload was requested on 2026-10-26.

The model repairs a draft video through its already-normalized Wan latents. It concatenates the 16-channel noisy state and 16-channel draft, and predicts a 16-channel flow. Inference is causal, one latent frame per chunk, with four SDE denoising steps and KV-cache updates. Text conditioning is the supplied archived empty context, with CFG 1.

These are custom Fixer/FastGen student weights, not a stock Diffusers pipeline. The exported FP32 tensors retain their exact checkpoint values and full transformer.* keys. Optimizers, fake-score model, discriminator, teacher weights, and training data are not included. This is an inference checkpoint, not a complete training-resume bundle.

Contents

• student.safetensors: complete student state dict.
• Wan2.1-T2V-1.3B-Diffusers/transformer/config.json: matching 32-input / 16-output architecture.
• empty_prompt_context.pt: exact conditioning used during training and evaluation.
• runtime/: matching Fixer adapter code and its local import dependency.
• load_student.py: strict loader and four-step latent repair helper.
• inference_config.json, training_summary.json, export_verification.json: settings and export checks.

Loading

Use the FastGen environment at commit c40fbea17b9a28a38528ee5254471dac4f2f9f12 from NVlabs/FastGen, with PyTorch 2.8.0, Diffusers 0.35.1, Transformers 4.49.0, and PEFT 0.17.1. Install the pinned FastGen dependencies and make its package available on PYTHONPATH.

import sys
from huggingface_hub import snapshot_download

folder = snapshot_download("primecai/fixer-concat-self-forcing")
sys.path.insert(0, folder)
from load_student import load_student, repair_latents

student = load_student(folder)  # CUDA / bfloat16 for inference
# draft: already-normalized Wan latents [B, 16, 21, 60, 104]
repaired = repair_latents(student, draft, folder, seed=20261025)

Do not apply Wan latent normalization a second time to cached training/evaluation latents. The original Wan VAE is a separate dependency for encoding/decoding RGB; it is not bundled here. The 21 latent frames correspond to the pilot's 81-frame, 832×480 videos.

Keep the supplied Wan2.1-T2V-1.3B-Diffusers directory name: the pinned FastGen loader recognizes the model family from that path. Loading is strict; the 32-channel concat projection must not be replaced by the stock 16-channel T2V projection.

Training and inference settings

Warm start: 1,000 causal flow-matching updates, LR 5e-5. Self-Forcing: DMD2 plus GAN, LR 5e-6, one student update per five outer SF iterations. This export is iteration 1,400 (280 SF student updates). Each run uses 32 H200 GPUs across four nodes, global batch 64, microbatch 2, and no gradient accumulation.

Inference uses four-step SDE sigmas [0.999, 0.937, 0.833, 0.624, 0.0], context noise 0, and one latent frame per causal chunk. The teacher initialization is the trained concat Fixer step-5873 teacher; this is not a VACE model or a depth-conditioned model.

The separately shared SF1,000 demo videos use an earlier checkpoint. This SF1,400 export has not been given a new qualitative evaluation. Training is still in progress.