CrossView-Warp lora workflow for Minimax H3

README.md

@@ -19,6 +19,14 @@ Drag-and-drop any `.json` file directly into ComfyUI to load the workflow.
 
 ---
 
+## MiniMax H3
+
+| Workflow | Description | Type | Updated |
+|---|---|---|---|
+| [CrossView Warp](https://huggingface.co/datasets/Cseti/ComfyUI-Workflows/blob/main/minimax-h3/crossview-warp/README.md) | V2V new-camera-angle LoRA — MoGe-2 geometry warp as an aligned guide, source clip as a Ref2VA reference, two-pass with audio (step 3.5k) | Video gen | 2026-09-20 |
+
+---
+
 ## LTX Video 2.3
 
 | Workflow                                                                                                                                                   | Description                                                                      | Type      | Updated    |

minimax-h3/crossview-warp/README.md

@@ -0,0 +1,47 @@
+# MiniMax-H3 CrossView Warp — Video-to-Video new camera angle
+
+Takes a video and a camera offset and regenerates the same scene from that new
+viewpoint, with audio. MoGe-2 estimates metric depth, the CrossView Warp node
+reprojects the clip into the requested camera pose, and a custom-trained LoRA on
+MiniMax-H3 fills the disocclusion holes and completes the view. The warp carries
+the geometry and the original clip carries identity and appearance, so the two are
+fed to the model in different roles: the warp as an aligned guide, the source as an
+ordinary Ref2VA reference. Two-pass, base generation followed by a refinement pass.
+
+## Preview
+
+![preview](media/preview.jpg)
+
+[preview.mp4](media/preview.mp4)
+
+## Requirements
+
+- ComfyUI with native MiniMax-H3 support (`MiniMax H3 Add Guide`, `MiniMax H3 Reference To Video`)
+- Custom node: [ComfyUI-CrossViewWarp](https://github.com/cseti007/ComfyUI-CrossViewWarp) — includes a walkthrough video for the camera controls
+- Custom node: [ComfyUI-KJNodes](https://github.com/kijai/ComfyUI-KJNodes) — resize, labels, image range
+- Models:
+  - [MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) — ref2va transformer, Qwen3-VL text encoder, video VAE, audio VAE
+  - [MiniMax-H3 CrossView Warp LoRA](https://huggingface.co/Cseti/MiniMax-H3_Ref2VA-LoRA-CrossView-Warp_v1)
+  - MiniMax-H3 DMD 8-step turbo LoRA
+  - MoGe-2 ViT-L (`moge_2_vitl_normal_fp16.safetensors`)
+
+## Notes
+
+- **Frame count must be 17k+5** — 73, 90, 107, 124, 141. Anything else is rejected or snapped.
+- **Resize the warp and the source to the same size** before the guide and reference nodes.
+- **Leave `pivot_override` off.** The node then estimates the orbit centre from MoGe's
+  metric geometry. Hand-entered `pivot_z` values are the fastest way to destroy the warp:
+  on a wide shot a wrong pivot gave 84% disocclusion with the subject gone, where auto
+  pivot at the same angle gave 52% with everything intact.
+- **Prompt is one word: `crossview`.** The LoRA was trained with that single caption and
+  no scene description.
+- LoRA strength 0.8, `res_multistep`, 8 steps with the turbo LoRA, sigma shift 12 video / 3 audio.
+- Angle range: azimuth up to +-90 is covered by the training data, elevation -20 to +30.
+  Looking up from below is the weak direction. Distance is not followed reliably, leave it at 1.0.
+- Wide shots with strong foreground are the hard case: if the subject is far and something
+  sits close to the lens, the camera move shreds the near field.
+- Output paths in the workflow are relative (`crossview_warp/...`), set them to your own.
+
+## Changelog
+
+- `2026-09-20` — initial release, LoRA step 3500

minimax-h3/crossview-warp/crossview-warp-h3.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:87882facdfee8a8a58b6abb82222530f45d11b32f951a33f95ae16fb9ae74994
+size 126336

minimax-h3/crossview-warp/media/preview.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd54acbec806dbc1fab4b805ff243d07d9e7c388b754cca0fc90f873d7940d26
+size 9978964