In the ComfyUI ecosystem, a production-like workflow has been demonstrated that separates control of a generative scene into three independent layers: actor movement, virtual camera, and final visual style. The chain connects Meta's open-source SAM 3D Body model, Blender, and ByteDance's Seedance 2.5 video model: first, regular video is converted into a full 3D body template, then in the 3D editor the mise-en-scène and camera are set separately, and only after that is the final scene generation launched. In such a scheme, control over movement and angle no longer depends on the accuracy of the text prompt.

image

What happened

The foundation of the chain was SAM 3D Body — an official ComfyUI template, available in the Template Library and on Comfy Cloud. Based on regular video, it reconstructs a full 3D model of the human body with pose, shape, and facial expressions, and its pipeline includes SAM3.1 multiplex tracking, RT-DETR v4 human detection on the DINOv3 backbone, optional MoGe 2 ViT-L camera angle estimation, facial expressions via MediaPipe, and export of a ready-made rig in GLB format. Next, the template is transferred to Blender, where the director independently of the original shooting changes the position of the virtual camera, angle, lens, and movement trajectory. The final scene is assembled by ByteDance Seedance 2.5 — a video model launched on July 31, 2026: it outputs up to 30 seconds of video in one pass with synchronized sound and lip-sync, accepting up to 30 reference images. Connecting Seedance 2.5 to the ComfyUI graph is done through partner nodes, meaning a closed API model is called directly from the workflow.

Context

None of the components of the chain is a new architecture or a new research result: all the novelty is in the orchestration and in the decomposition of control into three independent layers. Previously, the result of AI video was determined by a single text prompt, which made the staging of complex scenes a "lottery" of repeated attempts; the R2V scenario makes the deterministic 3D template the main control channel, and assigns a supporting role to the prompt. The logic itself repeats the classic scheme of 3D production preparation, where cheap previz is used to check the mise-en-scène and timing before expensive final rendering. At the same time, the mechanics of partner nodes have already become a standard way to connect closed commercial models to open-source ComfyUI graphs, so the described workflow relies on verifiable public blocks: official templates, docs.comfy.org documentation, and free local launch.

Why this matters for the industry

For the industry, the economics of AI video production are changing. The stage where the director sets the mise-en-scène, positions the camera, and checks the timing is performed for free on open-source tools SAM 3D Body and Blender, iterations are carried out in 3D before turning to the paid model, and money is spent only on the final passes of Seedance 2.5. This is a shift from the "pay for every attempt" principle to the "free iterations, pay for the result" model, which for the first time makes the classic previz stage part of an AI pipeline from publicly available blocks. For studios and builders working on advertising and short-form content, short scenes of up to 30 seconds per pass provide a working check of timing and camera before expensive generations. If the combination takes root, competition will unfold for the place of partner nodes in such pipelines, and other video model vendors will probably have to add an equivalent of R2V mode with 3D conditioning.

Why this matters for users

The chain can be reproduced immediately: ComfyUI and SAM 3D Body run locally and for free on your own video card, Blender is also free. The SAM 3D Body template is in the Template Library under the name "SAM 3D Body: 3D Human Mesh Extraction From Video", the same template is available on Comfy Cloud, and ready-made configurations are opened via cloud.comfy.org links. There is one paid stage: for Seedance 2.5, a Comfy account with credits is required, which are replenished by prepayment through Stripe, and the cost of a pass is shown before generation is launched. In practice, this means that the trial and error of poses, angles, and camera trajectories is carried out in the 3D template without costs, money is spent only on the final version, and the text prompt ceases to be the only way to control the scene. A realistic scenario for the reader today is prototypes and short scenes of up to 30 seconds, not long finished content.

What is still unknown / limitations

The sources do not confirm key launch characteristics: GPU requirements for local operation of SAM 3D Body are not specified, the time of a full pass is not measured, and the proportion of failed generations is not named. Quality assessment is absent as a class — there are no metrics for the accuracy of body, pose, and facial expression reconstruction, no measurements of drift between the 3D template and the final video, so the "production-likeness" of the workflow remains an interpretation of the demonstration, not a measured statement. The main capability statement — high correspondence of the final scene to the 3D template when generating Seedance 2.5 — requires verification on independent examples. The exact price of Seedance 2.5 generation is not disclosed in the available materials, and practical application is currently reasonably limited to the role of previz, not a guarantee of the final result.

Sources

Author

Look at AI, editorial team