A long-video workflow for MiniMax H3 Ref2VA built around the HR Endless Sampler, tuned and measured on a 12 GB card.
What it does: the latent is cut into chunks; before each chunk Gemma 4 12B looks at the previous frames and writes that chunk's prompt, and the last 22 frames are passed on as a video/audio reference for continuity. Your [Shot N] timings are respected.
Measured (RTX 4070 SUPER 12 GB, i5-12600K, 96 GB DDR5, Windows 11, ComfyUI 0.43 master, turbo 4 steps, 480×864):
- 15 s / 362 frames — 376–390 s.
- The same graph with 32 GB RAM took ~425 s; 64 GB+ keeps all models in RAM between chunks.
Settings that work on 12 GB: chunk_frames 124, video_continuation 22 (full), MiniMax Low VRAM Attention with 4 head chunks, gemma4_mtp off, cache_gemma_preproduction off. If a later chunk runs out of memory, try chunk_frames 90, then 56.
Tips:
- Write action for the whole length. A 6-second action stretched to 15 s makes Gemma invent shots.
- Faces in wide shots are soft at 480×864 (a face is ~80–90 px tall) — keep important faces in medium and close shots.
- No dialogue: voice locks need per-chunk denoise masks, which HR Endless doesn't support, so H3 would voice the lines itself.
- Known issue: colour drifts between chunks. In our runs average saturation rose ~40% from the first to the last second.
- Compared with the core "MiniMax H3 Context Windows" node: windows were 2× faster (190–204 s) with stable colour, but one prompt loops the action, and one prompt per window gives hard cuts. For one continuous shot with a storyline, HR Endless did better for us.
Needs: ComfyUI-HR-Endless-Sampler (hradec), ComfyUI-KJNodes; models from Comfy-Org/MiniMax-H3 (ref2va pruned int8, ref2v turbo 4-step LoRA, qwen3vl 32B NVFP4 text encoder, video VAE int8, audio VAE). Gemma 4 12B QAT is downloaded by the node on first run.