# Settings reference Everything below documents what a dial does, what the shipped default is, and what breaks if you move it. The shipped values are a working production configuration, not a theoretical best. Three of them are deliberately non-default: `audio_lock` off, `join_anchor_noise` 0.002 and `join_blend` on. Shipped defaults favour explicit identity: `ref2va` with voice self-anchoring and the bank on. The configuration that went through blind review was `fl2va`, which carries no reference rows, and switching to it is two changes (see *Checkpoint choice*). Both chain seamlessly, and both files are the same size - the difference is reference rows, not weight. --- ## The shipped recipe ``` checkpoint MiniMax-H3 ref2va (GGUF Q8_0 / Q5_1 / Q4_0) sampler euler scheduler beta57 (full workflow; needs RES4LYF) beta (CORE; stock ComfyUI) steps 14 frames/shot 362 (~15.1 s at 24 fps, the trained maximum) resolution 1280 x 736 (landscape) or 768 x 1344 (vertical) fps 24 ``` **On the scheduler.** `beta57` comes from RES4LYF, not stock ComfyUI. Measured on an identical seed, it scored 10/10 for lip-sync against 8/10 for stock `beta` ("slight synthetic stiff quality to the mouth"), with image quality, skin texture, artifacts and audio judged equal. The full workflow ships `beta57` and lists RES4LYF as required; CORE ships `beta` so it keeps its zero-third-party-pack promise. Switching is one widget either way. Both orientations sit under the model's pixel ceiling. Resolution **cannot change mid-chain** — every shot in one run must be the same size. --- ### Checkpoint choice `ref2va` (shipped) carries **reference rows** - the mechanism behind `voice_ref`, `self_anchor_voice` and the identity bank. Those three do nothing on `fl2va`, which has no reference rows; on `fl2va` they only cost tokens on every sampling step. `fl2va` still chains well - the voice is carried by the frame relay rather than pinned - and it is the configuration that went through blind review. Both files are **exactly the same size** at every quant level - fl2va and ref2va are byte-identical in bytes, so there is no VRAM or disk saving in choosing one. "Lighter" means fewer tokens per sampling step, because there are no reference rows riding along; it has never meant a smaller file. What you actually trade: * **`ref2va`** carries reference rows - voice anchoring, the identity bank, and reference images. It holds a supplied keyframe only *softly*. * **`fl2va`** has no reference rows at all, but it **lands on a supplied frame**. Measured against the same frame: **26.35 dB on fl2va versus 16.15 dB (turbo, 6 steps) and 16.81 dB (stock, 20 steps) on ref2va + keyframe.** The stock control rules out the sampler - the checkpoints genuinely differ. fl2va can also take a first *and* last frame and plan a camera move between them, which ref2va cannot do at all. Rule of thumb: **ref2va when identity or voice must persist, fl2va when a shot must start exactly where the last one ended.** ## Drift on long chains, and the dials that fight it Chained shots accrete detail. Each one conditions on the previous shot's own output, so invented texture compounds. Measured over **ten shots at 960x544** with all three of the following on: | control | shipped | what it does | | --- | --- | --- | | `memory_frames` | `0` | The big one. The bank's RECENT slots feed accreted output forward on top of the pin. At `0` the only reference is shot 1, which nothing has been added to yet. Luma drift 1.055 -> 1.022 per hop, chroma 1.086 -> 1.039, and the drift stops *accelerating* (4.2->6.7%/hop becomes 2.3->2.0%). Identity and framing held; framing actually improved. | | `master_normalize` | `luma+contrast` | Levels brightness **and** contrast of the finished chain to one target taken from shot 1. Matching only the mean masks a contrast ratchet while it compounds underneath. | | `pin_renorm` | `on` | Holds each pinned latent at shot 1's sigma. The pin's own sigma climbs every hop, and that inflated pin is what the next shot is handed. | Residual is about **1.02 per hop**. Not zero - long chains still drift, slowly. `chain_gain_control` is the other half: texture ratchets about 1.3x per join, so set it to `flatten` on chains longer than about five shots. `pin_noise` is small and scene-dependent (-1% to -2% per hop) and gets worse above `0.10`. It is not the fix it was described as in 2.1.5. ### Dials that do nothing under `context_pin` Verified in source rather than assumed. All three ship at `0` so nobody spends time tuning a dead control: * `join_anchor_noise` - noises KEYFRAME latents, and `context_pin` creates no join keyframes. * `handoff_release` - belongs to `continuity=latent_handoff`. * `bank_ref_noise` - reaches bank reference *images*, which measurement shows are not the carrier. ### Two-pass upscale Needs the H3 latent-upscaler pack, and is **not usable** with `continuity = 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 cannot preserve either. The node stops with an error naming the clash rather than quietly weakening the join. Available on `cut`, `seamless`, `seamless_tail`, `first_frame` and `flf_chain`. * `two_pass_upscale` `OFF` - each shot renders low-res through `pass1_fraction` of the steps, upscales in latent space, finishes at full res. * `upscale_factor` `1.5` - pass-1 size is size/factor, snapped to /32. * `pass1_fraction` `0.4` - verified clean. Past ~0.5 pass 2 cannot erase the upscale pattern and a ghost/moire lattice appears. * `upscale_audio_denoise` `0.35` - how much pass 2 may rewrite audio. `0` locks pass-1 audio (safest for voice), `1` is a full remix. --- ## MASTER CONTROLS (`H3StudioControls`) One panel drives `width`, `height`, `frames_per_shot` and `steps` on the sampler, and — in the full workflow — the prompt writer's dialogue pacing, so the LLM sizes lines to the real shot length. The panel also emits `sampler_name` and `scheduler` as strings. Those cannot link to the sampler's combo widgets on current ComfyUI frontends, so they go to the sampler's `sampler_override` / `scheduler_override` inputs instead: when connected they win, and when nothing is connected the sampler's own widgets apply. `shot_count` on the panel drives the sampler *and* the prompt writer's `num_shots`, so the two can never disagree. `0` means one shot per `---` block in the script, and lets the writer decide how many to write. `use_file_prompts` selects where the scene comes from: **off** reads the manual scene-idea box, **on** reads the prompt set (file or folder). The switch is lazy, so the branch you are not using never executes. --- ## Sampler dials ### Chaining | Dial | Default | What it does | |---|---|---| | `shot_count` | `0` | The TOTAL number of shots, **not** shots per prompt. `0` = one shot per `---` block in the script - leave it there for a written scene. `1..8` forces the total: extra blocks are dropped, a short script repeats its last block. | | `continuity` | `context_pin` (full) / n/a (CORE) | `context_pin` pins the previous shot's last 22 frames as **raw latents** — needs the Motion Context pack. `first_frame` uses the model's own trained hand-off — no extra pack. `cut` for episodic work. `seamless` and `seamless_tail` are **legacy** modes kept for comparison: `seamless` is a latent-only soft pin and often still reads as a cut; `seamless_tail` needs interior keyframe anchors and **conflicts with the Motion-Context pack** - with it installed the run stops up front with the alternatives named. | | `chain_gain_control` | `off` | Set to `flatten` for chains past about 5 shots. Each shot's tail anchors the next and the model returns ~1.3× the anchor's texture energy, so sharpness **ratchets** across a long chain with a visible step at every seam. `flatten` levels every shot to one house texture. | | `color_level` | `off` | Levels each shot's colour statistics to shot 1's settled tail. Not needed when chaining by latents — colour already carries. Useful if you see a warm/cool drift across a long chain. | ### Identity and voice | Dial | Default | What it does | |---|---|---| | `seed_per_shot` | `ON` | **Leave it on.** Measured: varying the seed per shot *holds* the face; using one seed for every shot drifted both face and voice. Identity lives in the conditioning, not the seed. | | `start_image` | unwired | An identity anchor image. Seeds shot 1 and anchors appearance. Optional — the frame relay plus verbatim descriptions usually suffice. | | `reference_images` | **off** | A batch of character portraits carried into **every** shot as ``, ``… Bind them in the prompt text. Needs a ref2va checkpoint. Fed by the REFERENCE IMAGES group two ways: **USE AUTO REFS** on (photos found from the script, nothing else to flip), or the two `LoadImage` slots with the **MANUAL REFS gate** on; chain another `ImageBatch` for a third and fourth manual slot. Unlike `start_image` these are not a first frame — they do not constrain shot 1's composition, they only carry who the person is, and they are what covers shot 1 while the memory bank is still empty. | | `voice_ref` | unwired | A clean solo speech clip, carried into every shot as `