Why this matters MiniMax‑H3‑Fun‑Controlnet‑Union consolidates multiple control modalities into one small control branch that you load on top of MiniMax‑H3. That design removes the need to swap per‑condition checkpoints when doing video‑to‑video or image+text→video generation, and its guidance‑distilled setup enables single‑pass inference with guidance_scale=1.0.
What Sets It Apart
- One checkpoint, multiple controls: a single safetensors file carries only the control branch (≈6.8 GB) and supports Canny, Depth, HED, MLSD and Pose controls as well as inpainting — no per‑condition checkpoint switching required. This simplifies pipelines that must handle heterogeneous control inputs.
- Lightweight integration model: the control branch attaches to five transformer blocks (layers 0,10,20,30,40) and uses zero‑gated projections to add control skips, with a configurable control_context_scale to tune guidance strength.
- Inference and resource notes: the checkpoint is guidance‑distilled (use guidance_scale=1.0), expands control input to control_in_dim=49 for inpainting (latent+masked_latent+mask), and follows the control video's frame/aspect constraints (frames snap to largest 17*n+5; duration ≤15s; 24 fps). Full pipeline requires the base MiniMax‑H3 checkpoint and significant memory; use group offload or CPU/QFloat8 strategies on an 80 GB GPU.
Who It's For and Trade-offs
Great fit if you need a single, drop‑in control branch to steer MiniMax‑H3 for many video control types (Canny/Depth/HED/MLSD/Pose) and want simpler runtime logic for V2V and inpainting. Look elsewhere if you need a full, standalone model (this checkpoint contains only the control branch), require per‑condition custom architectures, or must run on very small GPUs — the full stack (transformer + Qwen3‑VL text encoder) is memory‑heavy. Also check the MiniMax H3 Community License for territorial and acceptable‑use restrictions before production use.
