H3 Long Shot Studio

Music video made by this API https://www.youtube.com/watch?v=rPmv9dpM0P8

A browser studio for making long AI videos with MiniMax H3, shot by shot. It runs inside ComfyUI at /longshot. Build a film from short Shots, review and approve each one, and the Studio stitches them into one continuous take or a cut sequence. Approved work is never thrown away.

Generate

• Reference-to-video like the ComfyUI MiniMax H3 workflow: picture references for characters, outfits and places, plus audio references for character voices.
• Join each Shot to its neighbours with Continuous (flows on from the Shot before), Bridge (pinned at both ends), Context (a new angle, with the last frame before as a reference) or Cut (a fresh shot from the references).
• Frame anchors: first or last-frame images at any join.
• Render all renders every Shot in order as one flowing chain. Multi-take with seed control, a render queue across projects, and an automatic retry when ComfyUI runs out of memory.
• A song as reference audio (<Audio 1>) for singing Shots.

Edit

• A zoomable timeline with one block per Shot: switch takes instantly, loop the seam, approve or reroll any Shot in place while its neighbours keep their takes.
• Insert, bypass or remove Shots anywhere, including before Shot 1.
• Drop in real video clips as Shots, bridged into the generated footage or hard-cut.
• Take history for every Shot. Nothing samples again unless it has to.

Deliver

• Export approved clips, plain or upscaled with RTX Super Resolution.
• Upscale the final film in chunks with flat RAM, with an optional ProRes 422 HQ master.
• Export or import a whole project as one .zip.

1. Minor timeline UI adjustments for ease of use
2. Minor shot box UI adjustments for alignment

Several Studio tabs at once, each on its own project.

1. Each tab has its own queue and runs its own jobs; ComfyUI takes the runs one at a time, in the order they arrive (a tab whose run waits behind another tab's says so). Progress, previews and results go to the tab that queued them.
2. A duplicated tab gets its own identity (it no longer steals the other tab's progress).
3. One project, one tab: a project open in another tab can't be opened in a second one (two tabs would save over each other); the project list marks it "open in another tab". A new tab opens the last project that isn't open elsewhere.
4. A job queued in one tab for a project that's open in another tab moves to that tab when its turn comes, so the result lands where the project is open.
5. The queue panel also lists what the other tabs have queued (greyed). If a tab closes with jobs still waiting, another Studio tab picks them up.
6. Lip Sync function works now. See Audio tab for audio control for lip syncing.
7. Face Refine Added, refine selected clip then export them as approved clips.
8. Stackable timelines for overlapping clips with handles for easier editing.

• ComfyUI with native MiniMax H3 support, and the MiniMax H3 model files.
• Two node packs (the Studio names any that are missing):

ffmpeg is needed (on PATH, or pip install imageio-ffmpeg with ComfyUI's Python).

Optional, each with a Click to install button in the Studio where it's used:

For SageAttention, start ComfyUI with --use-sage-attention.

Needs ComfyUI 0.35.0 or later (native MiniMax H3). On an older ComfyUI the page opens but says to update, and nothing renders.

cd ComfyUI/custom_nodes
git clone https://github.com/r34vtraining/Longshot_Studio
git clone https://github.com/r34vtraining/H3_Prompt_Compiler
git clone https://github.com/r34vtraining/H3_Longshot

Skip any you already have. Then install H3 Long Shot's requirements.txt with ComfyUI's own Python:

Restart ComfyUI and open /longshot on ComfyUI's address:

• portable or git install: http://127.0.0.1:8188/longshot
• ComfyUI Desktop: http://127.0.0.1:8000/longshot (its port is in Settings → Server Config)

Updating: git pull in the Studio's and H3 Long Shot's folders, restart ComfyUI, then hard-refresh the page (Ctrl+Shift+R).

1. In Settings, check the model, text encoder, VAEs and Turbo LoRA (the Studio fills in the ones it recognises).
2. Drop reference images on Cast & Scenes and give each a label (<hero>, <street>) and a short description. Drop a voice clip there too, and link it in the character's editor.
3. Fill in Style & Sound, then write your Shots using those labels.
4. Press ▶ Render all, or render Shot by Shot. Review each one, then Approve or Reroll.

Every Shot has a chip that sets how it joins its neighbours. Click it to switch.

• Render all follows the chips: Continuous chains from the Shot before, Context splits the run there, Cut starts fresh.
• ▶ on one Shot pins it to the rendered Shots next to it (approved or under review); a Continuous neighbour still queued or rendering is waited for. Continuous between two rendered Shots reads Bridge and is pinned at both ends. The Shots around it keep their takes.

• A Shot pinned at both ends has to travel to a fixed end state: give it at least 5 s and write the arrival into its text ("…ends with her at the doorway").
• To keep drift from building on long takes, use ▶ Render odd Shots (each on its own), then Render all to bridge the gaps.
• Write speaker IDs like (S1) in the Shot text yourself; the Studio only adds the voice lines.
• Everything about a project (references, takes, videos) lives in input/longshot/<project> and output/longshot/<project>, and moves with Rename and Export.