Afflatus is a self-hosted studio that controls your own ComfyUI through forms, a media library, and timelines. For MiniMax H3, its useful distinction is clip-level creation with saved settings, alternate takes, and continuation through the model's latent state. It does not bundle the H3 weights or replace the GPU backend.
This guide follows the project README, usage guide, and ComfyUI setup documentation, checked October 16, 2026. These are documentation-based steps; we have not installed the studio or measured its joins. H3Vid does not execute Afflatus timelines.
Connect the studio to the renderer
Download the appropriate release from the author's repository, unpack it, and run its launcher. The documented local interface opens at http://127.0.0.1:8720. The project also offers Docker and source installation. Read the platform status before choosing a bundle: the README says some non-Windows bundles are built by automation but have not yet been started on real hardware.
In Settings, open ComfyUI and enter the backend address. For a backend on the same machine, the typical address is http://127.0.0.1:8188. Use Test to inspect the version, GPU, and nodes; that action does not launch a render. Also provide the folder containing ComfyUI's main.py so Afflatus can install its small node pack and inspect model headers. Restart ComfyUI after adding the pack.
For a remote backend, follow the project's separate network setup and supply a reachable Afflatus address for backend downloads. Do not assume that a local address works from another computer. A studio window opening on a Mac also does not demonstrate that H3 inference runs on that Mac; locate the actual ComfyUI GPU before diagnosing performance.
Prepare only the parts your workflow uses
Settings → Models lists ready and missing files. The studio recognizes model headers rather than relying entirely on filenames, which helps when files have been renamed. Its default families use Comfy-Org repackages, while individual presets may reference third-party fine-tunes.
The Afflatus node pack enables H3 latent continuation and related timeline operations. Optional features have additional dependencies: latent re-sampling uses the named H3 upscale pack, certain acceleration controls use DaSiWa or KJNodes, and other finishing methods need their own models. Read the missing-pack message for the selected operation instead of installing every optional family.
The application and its node pack also have different licenses: AGPL-3.0 for the studio and MIT for the pack. Model and preset conditions remain separate. Optional face-likeness scoring uses additional research-licensed weights and is not necessary for basic timelines. Keep the execution and model choices explicit in any shared setup.
Make one shot before building a sequence
Create a project, open the Studio form, select the H3 model, and enter a short scene. Use the model's supplied defaults first. A draft keeps the form configuration, while a render becomes a library item. This separation lets you revise a setup without confusing it with a completed take.
Open a result and reuse its settings with the same seed for a controlled change. Change the action, lighting, or another single parameter and compare the results. The studio's time estimate learns from your own earlier renders; it is not a transferable speed benchmark for someone else's card.
Keep the results you want. The documentation says new renders initially remain unkept, with age-based movement to the bin and later deletion controlled by Storage settings. Star an important take and check retention before relying on the library as the only copy of production work. Back up the project data and exported deliverables deliberately.
Continue a clip and choose its take
Open Timelines and drop a picture or video from the library onto a track. Each clip has its own generation form; an image can become its first frame, with optional additional pinned frames. Render several takes and select the one that should play.
Carry on links a clip to its predecessor. For H3, the documented join uses the model's latent state so motion and audio can continue across the boundary. This is a different operation from uploading an arbitrary MP4 to an online video extender. A local workflow must retain the compatible state that continuation expects.
The timeline reports join cleanliness and differences in brightness, sharpness, and pace. Treat these as selection aids. A numerically smoother join can still choose the wrong acting, dialogue, or camera movement. Watch the preceding and following shots together with sound before promoting a take. Mixing models may need a deliberate cut or transition rather than pretending the same latent path applies.
Finish and export are separate decisions
Finish brings clips to their final size using the selected restoration, upscale, or H3 latent re-sampling path. The guide says a finished take keeps its draft's timing. That is useful for editing, but it does not establish pixel identity or equal quality between finishing methods.
Export then assembles the chosen sequence. The documented web option creates a smaller delivery copy; other export routes include editable material for conventional video editors. Keep the full-quality version and the per-clip sources when the next stage requires color work, sound editing, or alternate takes.
Afflatus is a young, single-user studio, so begin with a short recoverable project and verify the exported file before moving a larger sequence. For related local state-management approaches, compare Project Suite and Continuum. Their saved-state contracts differ; a shared H3 backend does not make their project folders interchangeable.