ComfyUI's open-source SAM 3D Body pipeline gives AI filmmakers full control over character motion and camera before a single Seedance 2.5 generation runs

The biggest source of wasted iterations in AI video has never been the model quality. It has been the gap between what a filmmaker intends a character to do and what the model decides the character will do instead. A new open-source pipeline published through @ComfyUI addresses that gap directly, chaining SAM 3D Body motion extraction, a Blender camera rig and Seedance 2.5 generation into a single workflow that any filmmaker can run locally or on Comfy Cloud.

The result, demonstrated in output shared by creators including @NACHOS2D_ and @amnoartist, is AI video where the body motion and the camera path are authored, not guessed at. That is a meaningful shift in creative agency, and it arrives without a subscription fee or a waiting list.

What the pipeline actually does

The chain has four steps. First, SAM 3D Body extracts motion data from a reference recording, whether that is your own performance, stock footage or a pre-existing clip. Second, that motion data is brought into Blender, where you can retarget it to a character rig and adjust timing. Third, you set the camera, choosing angle, movement and framing using Blender's full suite of cinematography controls. Fourth, the result feeds into Seedance 2.5 as a controlled generation input.

Because SAM 3D Body is open-source, the motion extraction step runs locally. There is no API call, no per-frame cost and no third-party service holding your data. The Blender stage is equally offline. Only the Seedance 2.5 generation step consumes credits, and by that point you have already locked the performance and the shot, so you are generating against a known target rather than iterating blind.

Why this matters against the current alternative

The standard workflow for most AI filmmakers today is prompt-based iteration: describe the motion, generate, review, adjust the prompt, generate again. For a complex action beat, that can mean eight to fifteen generations before something usable appears. Each one costs credits and, more expensively, time.

Reference-image-to-video workflows improved things, but they still leave body motion and camera movement largely in the model's hands. ControlNet-style approaches exist for some models, but they typically require technical setup that puts them out of reach for creators who are not comfortable with node graphs.

This pipeline sits in ComfyUI, which has become the de facto environment for serious AI filmmakers who want repeatable, documented workflows. The node-based interface means the chain is inspectable, shareable and modifiable. A creator who builds a shot in this pipeline can hand the workflow file to a collaborator and have the same result reproduced exactly.

What to test first

The most immediate use case is action and dance sequences, the category where prompt-based generation is most unreliable. Per @ComfyUI, the pipeline gives more control over both body motion and camera movement, which makes it particularly well suited to choreographed beats where timing and framing are not decorative but structural.

For filmmakers new to Blender, the motion retargeting step will be the steepest part of the curve. Blender's rigging tools are powerful but not beginner-friendly. If you have never used Blender for character work, plan a session with the rigging before you try to push the full pipeline end-to-end.

For those already comfortable in Blender, the priority test is camera path fidelity: whether the camera movement you author in Blender survives the translation into Seedance 2.5's generation. That is the variable most likely to require refinement across early uses.

What was not said

The announcement does not specify which Seedance 2.5 variant the workflow targets, whether that is the standard model, ArtCraft or another flavour. It also does not detail the motion extraction accuracy across different recording conditions, lighting environments or clothing types, all of which affect how cleanly SAM 3D Body can read a performance. No benchmark figures are given, so creator reports like those from @NACHOS2D_ and @amnoartist are the current evidence base, and anecdotes are not benchmarks.

There is also no date or pricing announced for any future hosted version. The current path is self-hosted via ComfyUI or via Comfy Cloud, and cost will depend on your Seedance 2.5 generation volume.

The workflow shift

For working AI filmmakers, the practical value here is upstream control: the ability to make creative decisions in a tool designed for creative decisions, Blender, and then pass the output to an AI model that executes rather than improvises. That is a fundamentally different production model from prompt-iteration, and it is the direction the most technically capable creators have been moving for several months. This pipeline makes that direction accessible without requiring custom code.