MiniMax-H3 Multishot — chained shots, one master, with audio — v1.0 — v2.1 Bugs in my bugfixes

Models you need

checkpoint MiniMax-H3 ref2va (GGUF Q8_0 / Q5_1 / Q4_0) -> models/diffusion_models text encoder qwen3vl minimax_h3 (+ its -mmproj sidecar) -> models/text_encoders video VAE minimax_h3_video_vae -> models/vae audio VAE minimax_h3_audio_vae -> models/vae

GGUF quants: huggingface.co/joeygambino/MiniMax-H3-GGUF — Q8_0 for 32 GB, Q5_1 for 24 GB, Q4_0 below that.

GGUF encoder pairing. ComfyUI-GGUF matches the -mmproj vision sidecar to the encoder by filename, in the encoder's own folder. Rename either, or split them up, and it loads the encoder without its vision tower — which presents as the model ignoring your reference image. This pack's CLIP loader raises instead of continuing blind, uses the only mmproj beside the encoder when there is exactly one, and takes an mmproj_name widget so you can point at the file directly.

Which workflow

H3_Seamless_Chain_CORE — start here. The same seamless chaining with zero third-party packs . Type shots into the script box and queue.

H3_Seamless_Chain_v2 — everything: master controls, LLM writer, VRAM panel, identity and voice anchors, episode/batch prompt source, boundary plates, audio spine. Optional lanes are gated off by default.

H3_Keyframes — one clip, anchors at chosen frame positions, per-anchor condition strength.

The two things that stop people on the first run

1. The prompt writer needs a model you have pulled

The full workflow points at a local Ollama with model_name = qwen3:14b . If it is not pulled, the first queue stops immediately:

LLM API error 404: model 'qwen3:14b' not found

Fix: ollama pull qwen3:14b . Any OpenAI-compatible endpoint works — its URL in base_url , its exact tag in model_name . ollama list prints the tags you have, and it must match character for character.

No LLM at all? Set the master panel's use_file_prompts to manual entry, delete the writer, and feed your own script into the sampler's script input — one prompt per shot, separated by --- on its own line. CORE already works this way.

2. A local writer will fight the video model for the card

Turn on unload_model_after on the writer. It frees its own model from Ollama the moment the script is written. Without it the model stays resident for the server's default five minutes — your whole first shot. ComfyUI's own eviction cannot reach it, because Ollama is a separate process with its own allocator, and Ollama's OpenAI-compatible endpoint has no keep_alive field to ask with; the switch calls the native endpoint, which honours it. On under 32 GB, prefer a remote endpoint entirely.

Settings: start here, change nothing

checkpoint ref2va sampler euler continuity context_pin scheduler beta57 steps 14 fps 24 frames/shot 362 (~15.1s, the trained maximum) resolution 1280x736 landscape or 768x1344 vertical

Leave every VRAM switch off and the reserve at 0 , and try a render before touching any of it. The activation reserve measures each shape and conditioning payload as it renders and sizes the pool itself; it holds on 24 GB cards as well as 32 GB. A hand-set reserve overrides that measurement, so a number that suited one shape becomes wrong for the next. Those switches exist to dig out of a spill the console has already named, not for pre-emptive tuning.

Resolution cannot change mid-chain , and the mux must stay at 24 fps — other rates audibly shift voice accents. Dial-by-dial reference in SETTINGS.md .

Writing a script that chains cleanly

The previous shot's last ~1 second is replayed at the head of the next and discarded. Four rules follow, and breaking them is what produces mid-word chops and pose jumps:

Open holding. Every shot after the first opens in the previous shot's exact closing arrangement, with no dialogue for ~2 seconds. Give it real micro-motion — a breath, a weight shift — so it does not read as a freeze.

Land settled. Every shot ends with ~2 seconds of quiet, back in a stable arrangement, all dialogue finished.

Never split a line across shots. Dialogue plus 4 seconds of hold and settle must fit the shot length. If it does not fit, move the whole line to the next shot.

Repeat descriptions word-for-word. Character appearance and the room/light description, byte-identical in every shot. An unnamed light source gets reinvented per shot, and that is where colour drift starts.

The LLM writer applies these for you. Hand-written scripts must follow them — PROMPTING.md has a worked four-shot example, and example_script.txt is ready to paste.

How the chaining works

context_pin carries the previous shot's last 22 frames as raw latents — never decoded to pixels and re-encoded — placed at interior keyframe coordinates, with a timeline-placed audio reference alongside. The regenerated head is trimmed on decode. Colour, motion and voice cross the boundary as data rather than as a description.

Motion is the clearest case. Hand the next shot a single frame and it knows position but not velocity, so pace can reset at the boundary. Measured on a steady-pace walk: a single-frame anchor with no memory bank wobbled at the join; context_pin held it, and so did the memory bank on its own.

first_frame is the alternative — the model's own trained hand-off, no extra pack, and what CORE ships with. cut for episodic work.

Identity and voice

Nothing wired — the frame relay plus verbatim descriptions hold a face surprisingly well. A ~40 s two-character scene held both faces with no reference images at all.

self_anchor_voice (on) — shot 1's own rendered voice becomes the reference for every later shot. No file needed; write shot 1 with a clean solo line.

voice_ref — a clean solo speech clip, pinned across the whole chain including shot 1.

reference_images — character portraits carried into every shot as . Bind them in the prompt text.

seed_per_shot (leave on) — measured: varying the seed per shot holds the face; one seed for every shot drifted both face and voice. Identity lives in the conditioning, not the seed.

When something goes wrong

404, model not found — the writer's model is not pulled. See above.

A word clips at a join — the script put dialogue too close to a boundary. Move the whole line; do not split it.

Sharpening increases every shot — the texture ratchet. Set chain_gain_control to flatten ; worth it past about 5 shots.

Stalls at 0 steps, or runs several times slower than usual — a VRAM spill, the driver paging to system RAM instead of erroring. The console now names it. Raise the reserve, or drop resolution, frames, or reference payload.

Red or missing nodes — an optional pack is not installed. Delete those nodes, or use CORE.

GGUF architecture error — the pack teaches ComfyUI-GGUF the minimax_h3 architecture at startup. If it persists, run python apply_gguf_arch_patch.py from the pack folder once and restart.

Audio dulls on a very long chain — expected; restart the chain on a scene cut, where a fresh start costs nothing.

What changed in 2.1

The full workflow now works on a clean install. It referenced a prompt-source node that had never been published, and drove the writer through inputs the upstream writer pack does not have — so the boundary rules never reached the model. It rendered, and it rendered worse than it should, with no error to explain why. Both fixed: that node ships here as RiftPromptSource , and the rules are written into the workflow's own system prompt as well as carried by the writer pack in this zip.

The chaining sampler's anchor switches now do something. voice_ref , reference_images , self_anchor_voice , preview_first_shot , two_pass_upscale , reference_image_size and the sampler/scheduler overrides were drawn on the canvas but absent from the class, so ComfyUI stripped them before execution. All real now, render-verified.

unload_model_after on the writer, described above.

SHOT COUNT on the master panel drives the sampler and the writer together so they cannot disagree; prompt source switches between a manual scene box and a prompt set, lazily.

Node titles no longer name a checkpoint or a switch position — a title like H3 model (fl2va) is a lie the moment you change the model.

Two-pass upscale cannot be combined with context_pin or latent_handoff , or with an audio spine: those carry raw latents, or one locked denoise trajectory, across the join, and a two-pass render preserves neither. The node stops with an error naming the conflict rather than quietly producing a weaker join. Two-pass is available on cut , seamless , seamless_tail , first_frame and flf_chain .

Credits

Prompt writer: RealRebelAI ( ComfyUI_JoyAI_Echo_GGUF_Nodes , modified — see the NOTICE in the zip). context_pin : NikoDemon80 ( ComfyUI-H3-Motion-Context ). Two-pass upscaling: Tr1dae ( ComfyUI-MiniMaxH3_LatentUpscaler ).

Text encoder · Unknown