# H3 generation rules (from Run 1 + Back Nine drafts, Nov 7 2026)

The do's and don'ts for running MiniMax H3 in the factory pipeline. Every rule comes from something that went right or wrong on Nov 7 (about 100 clips on 4x L40S). Items marked **UNTESTED** still need a test before they become rules.

---

## 1. Mode: r2v only

**Names (decided by Aydan, Nov 12):** a **keyframe** is one shot's layout drawn in the real look with the real cast, fed to H3 as the frame to open on. **Storyboard frames** are the loose ink drawings. **Sheets** are the locked character, place and prop images. **Lock frames** are the stills cut from a picked take for the final render. Older text below that says "storyboard / keyframe image" means storyboard frames.


- **DO** use r2v (reference-to-video) for every shot. Feed it the locked character sheets, location sheets and prop sheets. Up to 9 references, referred to in the prompt as `<Picture 1>`, `<Picture 2>`, and so on, in the order they are passed.
- **DON'T** use i2v (start frame / end frame). Aydan's verdict on Run 1: the i2v clips are bad. It also errored out at the largest size.
- ~~DON'T make keyframes (still images) before drafting.~~ **REPLACED Nov 12 (Aydan):** keyframes are a standard stage. Start keyframe only, never an end one (T8). The first shot and every cut to a new angle get a keyframe; a shot that continues the same camera starts from the last frame of the previous picked take. Image spend still needs a go.
- **DO** draw every storyboard / keyframe image in one style: **S3 ink + grey tone** (clean black ink lines, flat grey marker tones, high contrast, no colour). Locked by Aydan Nov 12 after a 6-style test on shots 1B and 2D (sheet: `work/test_run_2/storyboard_styles/`). Make them on the factory image worker with `api_model: gpt-image-2` (low quality, the cheapest usable). Never on Sume. Never redo one just because it is imperfect: success is the video lining up with the frame.
  - Style line to paste: "Comic-style storyboard panel: clean black ink lines with flat grey marker tones, high contrast, no colour."
  - **LOOSE, not detailed (Aydan, Nov 12, T8: "the loose one is way better").** Storyboards are for light stuff: people are generic gesture figures with blank heads (no faces, hair or clothing detail), props and buildings are simple blocks, background barely suggested. They carry only camera framing, where each person and object sits, and poses. Detailed storyboards got copied exactly by the stills, mistakes included. Draw them from text (no sheet refs), but write the layout out explicitly (who sits where, which side, how many of each prop), or it drifts. Identity and look come from the sheets at the stills step. Prompts: `work/test_run_2/t8_chain/loose/make_loose.py` + `make_loose2.py`.
  - **T8 verdict (Aydan, Nov 12): best = loose storyboard as a reference (V2L) or a start still made from the loose storyboard (V3L).** Both beat sheets-only (V1) and every detailed-storyboard arm. **DON'T add an end still (V4):** H3 does weird, cliche moves to force its way to the end frame. Start only. Comparisons: `work/test_run_2/t8_chain/labeled/T8_1B_compare.mp4`, `T8_2D_compare.mp4`.
- **UNTESTED:** t2v (prompt only). It has no place yet: with no references, characters drift.

## 2. References and continuity

- **DO** lock every character, location and prop as a sheet before any video. One character per sheet. A sheet that shows two people bleeds both into the shot (Mike and Priya had to be split).
- **DO** pass only the references a shot needs. Each extra reference makes the clip slower (2 refs: about 580 s for 10 s; 4 refs: about 630 s).
- **DO** carry continuity forward. Once a shot is picked, take the last frame of the picked take and pass it as an extra reference to the next shots in the same scene. In the prompt, add: "<Picture N> is a still from the previous shot: keep the same table, seating positions, clothes and lighting." This fixed the seating jumping between 1A, 1B and 1C.
- **DON'T** draft shots in the same scene in parallel before the first one is picked if seating or layout matters. Pick shot A, then make B and C from A.

## 3. Writing the prompt

- **DO** open with the same look header on every shot of a brand (for Back Nine: the PS2 game look line), then one line naming each `<Picture N>`.
- **DO** write the action as timed beats: "At 00:01.000, ...". One action per beat.
- **DO** split any hand action into separate beats: hand reaches, fingers close, then the object moves. If you don't, objects float before the hand touches them (the cup in Run 1).
- **DO** state where every object IS, in every beat it matters. H3 has no negative prompt, so "don't" never works. Example: "the ball rests flat on the grass two metres from the hole" fixed the ball-on-a-tee-in-the-hole problem. "The putter stays in his left hand" fixed the missing putter.
- **DO** state physical direction: which way a door hinges and swings, which side of the cart he sits on, which hand holds what.
- **DO** describe signs and screens as "plain blank" so there is no gibberish text.
- **DO** make the scene logic hold up: if a cart is moving, someone is driving it. If nobody is in it, it is parked. An empty seat on a moving cart reads as a ghost car (2C).
- **DO** put spoken lines in `<d>[English] ...</d>` with speaker labels (S1), (S2), (S3), each line on its own timed beat.
- **DO (LOCKED Nov 11, T1): voice comes before video.** Record the line in ElevenLabs first, approve it, then pin it as the shot's soundtrack from frame 0 (MiniMaxH3AddGuide). H3 animates the lips to it. Method B: 4 of 4 seeds said the line once, on time (checked with faster_whisper). Passing the audio as a loose reference ("Use <Audio 1> exactly as it is") failed 4 of 4: started early and invented words. Still unproven for two speakers in one shot (G1A never picked).
- **DON'T** ask H3 for off-screen narration, voice-over or music. They come in the edit. If H3 voices narration, mouths flap.
- **DON'T** ask for on-screen text, captions or logos inside the video. Add them in the edit.
- **DON'T** let a cutaway repeat another clip's scene. Check the storyboard for overlap before writing.

## 4. Settings

| Use | Size | Steps | Time per 5 s clip (one L40S) |
|---|---|---|---|
| Drafts (pick performances) | 480x864 (0.4 MP) | 20, turbo off | about 4.5 to 5 min |
| Final size, if rendered natively | 768x1344 | 20 | about 13 to 14 min |

- **STANDARD (Nov 7): every shot that goes in a final video is rendered at 768x1344.** Drafts at 480x864 are only for picking performances. A 480x864 face in a wide shot is about 20-30 pixels tall, and no upscaler can rebuild it (see section 7).
- **DO** draft at 480x864, full 20 steps, 4 seeds per shot.
- **DO** lock the final render to the picked draft: same prompt, same seed, plus 2 stills from the picked take (opening frame + middle frame, resized to 768x1344) as extra references, with this line after the picture list: "<Picture N> and <Picture N+1> are stills from the approved take of this exact shot (its opening frame and its middle frame). Recreate that take: the same camera framing and angle, the same character positions and poses, the same props in the same places, the same lighting, and the same action at the same moments." One render per shot, no new seeds to pick through. **CONFIRMED (Nov 7):** 1A and 3A fix came back almost identical to the picks (same layout, poses, timing), only small drift.
- **DON'T** expect a seed to carry across sizes. The seed sets the starting noise, and the noise is a different shape at a different size, so the same seed at 768x1344 is a different take. Same seed + same size + same prompt = the same clip.
- **DO** frame characters medium or close where faces matter. Wide shots with small faces are where quality falls apart.
- 10 s at 768x1344 with 6 references fits on an L40S (about 45 of 46 GB in use, tight) and takes about 37 min (about 106 s per step x 20).
- **DO** make 10 s shots only when the action needs it. They take about twice as long as 5 s.
- **DON'T** use 720x1280. It is not a valid H3 size.
- **DON'T** expect the same seed at a different size to give the same take. A new size is a new take, so never re-render a pick bigger to "upscale" it (see section 7).
- CFG stays at 1.
- **Reference video feed size is a choice, not a fixed setting (Aydan, Nov 12).** Applies to shots that copy motion from real footage (T3 fix 2 style picture / fix 4 outline, `--ref-video`). Final render is 768x1344 either way; only the size the real clip is fed in at changes (`--ref-video-size`).
  - **Big feed** (reference at 768x1344, the default): about 39 min per 5 s clip. Keeps fine detail from the real clip (cut 8b: the dirt off the ball reads clearly).
  - **Small feed** (`--ref-video-size 480x864`): about 23 min per 5 s clip (cut 8b style: 1375 s vs 2327 s). Same swing, pose and framing; fine detail slightly softer. Not a huge difference.
  - **Studio UI:** the motion-copy stage offers both as a selection, with a warning showing the time difference (about 16 extra min per 5 s clip for big feed) before the job is queued.
  - **Stage sub-agent standard:** the agent that owns this stage asks for (or reads) the feed choice per shot, never picks silently, and states the time cost when it proposes the job. Comparison: `work/test_run_2/t3_out/labeled/CUT8b_fix2_STYLE_smallfeed_compare.mp4`.
- **PARTLY TESTED (Nov 11, T5):** turbo 4-step about 88 s, 8-step about 132 s per 5 s clip. Aydan: "worse quality but similar video in general". Still unproven: whether a turbo pick locks to the same take at 768x1344. Test that before using turbo for drafts.

## 5. Running the box

- **DO** pull the public H3 weights from Hugging Face (free, about 230 MB/s). Keep B2 for our own files: 42 GB from B2 used about 75% of the daily download allowance.
- **DO** run one ComfyUI copy per card (ports 8188 to 8191) and start all 4 cards at once. Do not wait on the first clip.
- **DO** feed the cards from one queue file that every card pulls from (`h3_queue.py`). It re-reads the file for each clip, so clips can be added or reordered while it runs. Use that to push fixes to the front.
- **DO** set a hard auto-terminate and spend cap when renting.
- **DON'T** run anything else on a card while H3 is on it. H3 at draft size uses most of a 46 GB card, and even a tiny test can fail with out-of-memory.
- **DO** unload H3 from ComfyUI before running anything else on the cards, even after the queue is empty. Idle ComfyUI keeps about 39 GB per card loaded. Run `curl -X POST localhost:<port>/free -d '{"unload_models":true,"free_memory":true}'` on ports 8188 to 8191 first (the first upscale try on Nov 7 failed out-of-memory because of this).
- **DON'T** launch background jobs on the box through a plain `nohup ... &` over ssh. Use `setsid nohup ... & disown`, then check that exactly one copy is running.
- Throughput reality: 4 cards gives about 50 draft takes (5 s) an hour. A 14-shot storyboard at 4 seeds is about 2 hours. The bottleneck is GPU compute (cards verified at 100%), not setup.

## 6. Review loop

- **DO** post each shot's 4 takes to one review page as soon as all 4 finish, in storyboard order.
- **DO** log every pick and every complaint in the run log with the reason. Each complaint becomes a prompt fix (section 3) and a retake row next to the original.
- **DO** retake only the shot that failed, with a fixed prompt, 4 seeds, at the front of the queue.
- **DO** check whether a flip solves it before re-rendering. Mirroring a take left to right is free (2B seed 2). Watch for text or logos that read backwards.
- **DO** stitch a rough cut as soon as every shot in a clip has a pick, to judge flow and continuity together.
- **DO** check the first second of every locked final. A shot that carries a continuity still from another shot can open with a half-second flash of that other shot (1C did). Trim it off; it costs nothing.

## 7. Finishing

- **DO** upscale only the picked takes, after picking, never the drafts.
- **DON'T** rely on upscaling to rescue 480x864 drafts. Tested Nov 7 on all of Clip 1, Aydan's verdict "not looking good":
  - SeedVR2 7B (free, on the box, about 5 min per 5 s clip): sharpest, but invents detail on small faces (the face went weird in the X2 tee stretch).
  - SeedVR2 3B (free, on the box): same problem, less detail.
  - Sume video upscale, standard tier ($0.0069 per second, $0.25 for 36 s): softer, still not good on far shots.
  - None of them take a prompt. Prompted tools that "restyle" video re-generate it and change faces and motion, so they undo the picking.
- Upscaling is only worth a test on footage that is already native 768x1344 (for example to 1080 for delivery), never as the quality fix.
- Audio: on-screen dialogue is the ElevenLabs line already pinned into the shot (section 3, locked Nov 11). Off-screen narration and music go in the edit.

## 8. Never

- Never start a GPU or image spend without a go from Aydan.
- Never use i2v for production shots (section 1).
- Never write "no X" in a prompt and expect it to work. Describe what is there instead.
- Never put two people on one character sheet.
- Never leave the box idle while waiting for a review: keep the queue full with the next shots.