Nodes/ComfyUI-H3-Multishot/H3 Multishot Sampler + Memory (long form)

ComfyUI Node

H3 Multishot Sampler + Memory (long form)

Long-form chains with a bank that holds the story

By jlucasmcrell·Created 21 days ago·Updated 2 days ago· 31

model
clip
video_vae
audio_vae
start_image
keyframe_images
guide_audio
reference_images
voice_ref
sigmas
reference_video
reference_video_audio

master_frames
master_audio
shots_rendered
video_latents
audio_latents
head_frames
master_path

◄script►

◄shot_count0►

◄width768►

◄height1344►

◄frames_per_shot243►

◄seed0►

◄steps20►

◄seed_per_shottrue►

◄memory_frames0►

◄anchor_frames1►

◄sampler_nameres_multistep►

◄schedulersimple►

◄bank_pinned1►

◄chain_gain_controloff►

◄continuitycut►

◄bank_clip_frames22►

◄color_leveloff►

◄join_anchor_noise0.000►

◄join_blendfalse►

◄handoff_release0.30►

◄bank_ref_noise0.000►

◄end_anchorfalse►

◄join_fxoff►

◄audio_locktrue►

◄handoff_taper0►

◄handoff_depthblock►

◄sampler_override—►

◄scheduler_override—►

◄self_anchor_voicefalse►

◄reference_image_sizematch►

◄preview_first_shotfalse►

◄output_scale1.00►

◄save_every_shotfalse►

◄upscale_model_name(none)►

◄master_normalizeluma+contrast►

◄pin_frames22►

◄pin_noise0.050►

◄pin_renormtrue►

◄reference_subjects►

◄low_ram_masterfalse►

◄audio_pin_frames0►

◄pin_noise_audiofalse►

◄audio_tone_controlfalse►

◄x0_texture_clamp0.000►

This is the flagship. Where the classic H3MultishotSampler relays a single frame between shots, the memory sampler adds an identity bank and a choice of join mechanisms - including context_pin, the tightest join this pack has, which carries the previous shot's last 22 frames into the next as raw latents with no VAE round trip. It's the node behind the shipped H3_Seamless_Chain and H3_Extend_Take workflows, and it's the one you want for anything longer than a couple of shots where identity and voice have to persist.

How the chaining works

The memory bank is the core idea. It keeps reference clips of shots in the latent domain, so later shots always have something to hold onto. Two levers control it: memory_frames (how many most recent shots stay in the bank, default 0 - deliberately, because recent slots hand each shot's accreted output forward as refs and invented detail compounds; measured texture growth dropped from 1.055 to 1.022 per hop moving 2→0) and bank_pinned (how many of the earliest shots stay permanently - the anti-drift lever, since with shot 1 pinned, later shots always see where the episode started).

Then continuity picks the join mechanism, and it's worth reading the tooltip in full because each mode is a genuinely different technique:

context_pin (the shipped default) - 22 raw-latent frames pinned into the next shot's head at interior keyframe coordinates, plus a timeline-placed audio ref. Bit-identical content, velocity and colour carry, regenerated head trimmed on decode. Needs the ComfyUI-H3-Motion-Context pack.
first_frame - the frame relay, no dependencies, uses fl2va's continuation task.
latent_handoff - one denoise trajectory; the next shot's first latent block is hard-locked to the previous tail at every step. Speech can continue mid-word.
flf_chain - true first/last-frame mode: N+1 boundary stills for N shots, every join one shared picture. Fails loudly without plates.
cut - JoyEcho-pure, no keyframe, a fresh take per shot held together by the bank (correct for intentional cuts).
seamless / seamless_tail - legacy soft pins, kept for comparison.

The inputs that matter

Beyond model/clip/VAEs and script (same --- format), the ones you'll actually touch:

seed_per_shot - leave ON; one seed across shots drifts face and voice.
start_image - not a first frame. It's an identity reference row (and does nothing on fl2va). For a real I2V start, use the classic sampler or flf_chain with keyframe_images. This naming trap has burned people.
guide_audio - the audio spine: a continuous track for the whole take, each shot's audio locked to its time-slice at every step. This is the locked-audio music-video path, works with every continuity mode, and resamples any sample rate to the audio VAE's.
voice_ref / self_anchor_voice - external or self-pinned voice identity (ref2va).
chain_gain_control - flatten on chains past ~5 shots; texture ratchets ~1.3x per join.
master_normalize - luma+contrast (default) deflickers the finished chain to one global target, outside the feedback loop.
reference_subjects - comma counts in picture order (3,3 = two people, three photos each) so multi-character refs don't average into one face.
save_every_shot and low_ram_master - insurance: save each shot as it decodes, or stream the master to disk so peak RAM is one shot.

Outputs: master_frames, master_audio, shots_rendered, plus raw video_latents / audio_latents / head_frames for advanced use (they're untrimmed - later shots open with replayed head that's removed at decode, so they don't line up with the master until you trim head_frames yourself) and master_path when streaming to disk.

Install and reality check

cd ComfyUI/custom_nodes
git clone https://github.com/jlucasmcrell/ComfyUI-H3-Multishot

or ComfyUI-Manager → H3 Multishot. ComfyUI v0.30.0+; the full workflow also wants ComfyUI-H3-Motion-Context (for context_pin), RES4LYF (the beta57 scheduler), and the writer pack - but the CORE workflow needs none of them, and every one can be removed with a documented one-widget change.

Two things the README hammers that you should internalize. First, the VRAM spill trap: shot 1 succeeding tells you nothing about shot 2 - shot 2 carries the pin and reference rows, so it wants roughly double the pool; if it won't fit, lower frames or resolution. Second, the honest limit list: very long chains still dull the audio per hop, and the ref2va + bank + context_pin combination hasn't been blind-reviewed as a unit (the reviewed recipe is fl2va). At ~4 windows it's excellent; at 7+ the texture ratchet gets visible. It's the most capable single node in the pack - and the one with the most dials, so start from a shipped workflow rather than from scratch.

Categorysampling/minimax

Inputs (56)

Name	Type	Default	Description
model	MODEL		—
clip	CLIP		—
video_vae	VAE		—
audio_vae	VAE		—
script	STRING		One prompt per shot: JSON {"prompts": [...]} or plain blocks separated by --- lines.
shot_count	INT	00–64	The TOTAL number of shots - not shots per prompt. 0 = one shot per --- block in the script (leave it here). Above 0 forces the total: extra blocks are dropped, a short script repeats its last block.
width	INT	76832–4096	—
height	INT	134432–4096	—
frames_per_shot	INT	2435–1450	Trained range is ~124-362; longer single shots are ladder territory (RoPE extrapolation) - the audio-spine pass runs 719f low-res through exactly this.
seed	INT	00–18446744073709550000	—
steps	INT	201–50	—
seed_per_shot	BOOLEAN	true	—
memory_frames	INT	00–3	RECENCY slots: how many of the most RECENT shots stay in the bank, on top of the pinned one. Total bank = pinned + recent, capped at 3 by H3's reference limit. DEFAULT 0, deliberately. The recent slots hand each shot's ACCRETED output forward as reference images, on top of the latent pin, so invented detail compounds. Measured over ten shots at 960x544, moving 2 -> 0: texture growth 1.055 -> 1.022 per hop, chroma 1.086 -> 1.039, framing correlation at shot 10 0.976 -> 0.995, and the drift stops ACCELERATING - 4.2%->6.7% per hop became 2.3%->2.0%. At 0 the only reference is shot 1, which nothing has been added to yet. Identity and framing held on a static scene. If a busy scene loses motion continuity between shots, raise it to 1.
anchor_frames	INT	10–9	Identity reference images taken from start_image, used on EVERY shot (JoyEcho seeds identity on shot 1 and lets the bank carry it after that; keep this at 0 or 1).
start_imageopt	IMAGE		Optional identity reference image. NOT a first frame - this node has no keyframe.
keyframe_imagesopt	IMAGE		flf_chain only: N+1 boundary stills for N shots, in order. Shot i is generated between image i and image i+1. Best source is a single long low-res take of the whole scene - its frames at the boundary times are already colour-matched, identity-matched and correctly posed, so every join inherits one consistent look.
guide_audioopt	AUDIO		AUDIO SPINE: a continuous audio track for the WHOLE take - a voice recording, a song, or a low-res long pass. Each shot's audio stream is locked to its time-slice of this track at every sampling step; the video follows the locked audio (lips included) via the model's own audio-video attention. Works with EVERY continuity mode - the per-shot stride accounts for each mode's seam trim (render-verified on context_pin). Any sample rate: the track is resampled to the audio VAE's rate and mono is upmixed. This is the locked-audio music-video path.
sampler_nameopt	COMBO	res_multistep	44 options: euler, euler_cfg_pp, euler_ancestral, euler_ancestral_cfg_pp, heun, heunpp2, +38
scheduleropt	COMBO	simple	9 options: simple, sgm_uniform, karras, exponential, ddim_uniform, beta, +3
bank_pinnedopt	INT	10–3	How many of the EARLIEST shots stay in the bank permanently. This is the anti-drift lever: with shot 1 pinned, later shots always see where the episode started. 0 = pure recency.
chain_gain_controlopt	COMBO	off	Texture levelling across the chain. flatten = level the DECODED frames and the bank to shot 1 (the pin still carries accreted texture; measured x1.13-1.15 per hop at 736x1280 with flatten on). flatten_pin = flatten PLUS a latent high-pass trim on the pin; leaves ~x1.058/hop residue. refresh_pin = the pin's video tail is decoded, levelled to house/(learned hop gain) with the calibrated flatten, re-encoded and spliced; audio rides raw (A/B 2026-08-23: joins passed blind review, drift x1.043 vs flatten_pin's x1.105 over 2 joins). level_pin = EXPERIMENTAL: closed-loop LATENT leveling - measure decoded texture, tilt the raw pin latents' high band, re-decode to verify, iterate; no VAE re-encode, so the pin stays native latent statistics.
continuityopt	COMBO	cut	cut = JoyEcho-pure: no keyframe, every shot is a fresh take held together by the bank. Framing and exposure step between shots - correct for multishot storytelling with cuts. seamless = LEGACY, kept for comparison: hands the next shot its predecessor's last frame as a latent-only keyframe. That is a SOFT hint - no vision tokens - and the model often satisfies it loosely, so the join can still read as a cut. For a real join use context_pin or first_frame. seamless_tail = LEGACY, kept for comparison: pins the previous shot's frames -9/-5/-1 at keyframe indices 0/4/8 so velocity carries too. Needs interior keyframe anchors, which CONFLICT with the Motion-Context pack - with that pack installed this mode stops with an error naming the alternatives instead of crashing mid-chain. latent_handoff = one denoise trajectory: the next shot's first latent block (video AND audio) is hard-locked to the previous shot's actual tail latents at every sampling step, released only for the final detail steps. Speech continues mid-word because the model wakes up inside its own previous state - no keyframes involved. first_frame = the model's OWN continuation mechanism (fl2va task): the previous shot's last frame is handed over the way the stock Image-to-Video node does it - as VISION TOKENS through the text encoder AND as the frame-0 keyframe latent. The new shot literally starts on that frame; only the duplicate is trimmed. USE AN fl2va CHECKPOINT (ref2va is trained for reference rows, not first-frame hand-off) - and note the bank is disabled here, because fl2va has no reference rows. flf_chain = TRUE FFLF. Supply N+1 boundary keyframes for N shots; shot i renders BETWEEN keyframe i and keyframe i+1. Shot i ends on exactly the image shot i+1 begins on, so the join is one shared picture rather than two independent guesses - colour, framing and pose match by construction. context_pin = Motion-Context chaining (needs the ComfyUI-H3-Motion-Context pack): the previous shot's last 22 frames are pinned into the next shot's head AS RAW LATENTS at interior keyframe coordinates - bit-identical content, no VAE round trip, so velocity AND colour carry - plus a timeline-placed audio ref. The regenerated head is trimmed on decode. Composes with the bank, colour levels and join fx.
bank_clip_framesopt	INT	225–124	Frames per bank slot, taken from the middle of each shot (JoyEcho stores a clip, not a single frame). Reference rows cost time on every sampling step, so keep this small: 22 is ~0.9s.
color_levelopt	COMBO	off	Colour drift across a chain. 'scene' is the one to use: ONE reference for the whole piece, applied per frame at the very end, so every shot is pulled to the same target and there is no step at any join. 'mvgd' is DEPRECATED - it matches each shot to a rolling house and leaves a hard step at every seam (measured 29% warmth step; a render with it on drifted +18% brighter over three shots). It is kept only so saved graphs still load.
join_anchor_noiseopt	FLOAT	0.0000–0.05	Mix this much seeded noise into every join keyframe latent (SkyReels noised-clean-condition). The texture ratchet exists because the model treats its own output as pristine and adds ~1.2x detail on top; a little noise closes that gap at the source. 0.02 is the researched setting. The noised frames never reach the final cut.
join_blendopt	BOOLEAN	false	seamless_tail only: instead of hard-dropping the 9 regenerated overlap frames, crossfade them against the previous tail (with a grain guard so the blend band does not read as a grain dip) and fade audio over the same 375ms. Any residual step is spread across 9 frames instead of landing on one boundary.
handoff_releaseopt	FLOAT	0.300–1	latent_handoff only: sigma below which the locked overlap is released so the detail steps can reconcile it with the new content. Higher = freer (released earlier), 0 = locked to the very end.
bank_ref_noiseopt	FLOAT	0.0000–0.05	Mix this much seeded noise into every bank clip before it is stored (SkyReels noised-clean-condition, same idea as join_anchor_noise but for the references). The texture ratchet rides bank clips: the model copies its own output and adds ~1.2x detail, worst on faces/skin. 0.02 is the keyframe-researched setting.
end_anchoropt	BOOLEAN	false	Pin shot 1's FIRST frame as a keyframe at the LAST frame of every later shot. H3 push-in creep compounds across chained shots (each shot inherits the previous crept tail and pushes further), so tails drift ever further from the prompted framing and every join mechanism inherits an off-spec tail. The end pin closes the loop: a shot may breathe inward mid-take but must settle back to house framing by its tail, so the next join starts from framing the text agrees with.
join_fxopt	COMBO	off	Diegetic join masking: dress every join in a short VHS tracking hiccup (displacement bands, chroma shift, dropout flecks, audio head-switch duck+hiss) peaking on the boundary. For analog-horror content the cut stops being an artifact to hide and becomes part of the tape. Works with every continuity mode.
audio_lockopt	BOOLEAN	true	latent_handoff only. ON: the audio head is a locked replay of the previous tail (or the spine). OFF = the SILENT-JOIN policy: script each shot to land its line and hold still for the last beat; audio generates freely, the new head's audio is kept in full and the previous tail's audio is trimmed instead. Nothing the model generates is discarded, so no word can be lost BY CONSTRUCTION. A connected guide_audio spine overrides this to locked.
handoff_taperopt	INT	00–10	latent_handoff only. Rows AFTER the hard lock that are softly biased toward a continuation of the previous motion, at linearly decaying strength. Without it the lock ends at a cliff: the replayed frames are faithful, then the next frame follows the shot's OWN pose plan - measured as a 3x larger discontinuity on the person than on the static room. The taper gives the pose a ramp to follow. 3-5 is a good start (each row = 4 frames).
handoff_depthopt	COMBO	block	latent_handoff overlap depth. block = lock the first full 17-frame block (strong video anchor, 22 frames trimmed, ~0.92s join gap). bootstrap = lock only the 2 bootstrap rows to the previous last 5 frames (5 frames trimmed, ~0.21s gap - an ordinary breath), with a frame-0 keyframe pin of the previous last frame; end_anchor + the bank carry the rest.
reference_imagesopt	IMAGE		Optional SUBJECT/CHARACTER reference images (batch = multiple refs, e.g. via Batch Images), carried into EVERY shot as <Picture 1>, <Picture 2>, ... ahead of the bank slots, so their numbering never shifts as the bank fills. Distinct from start_image, which seeds identity for shot 1 only. Bind them in each shot's prompt: 'She looks like the woman in <Picture 1>.'
voice_refopt	AUDIO		Optional VOICE ANCHOR carried into EVERY shot as a reference audio (<Audio 1>). Feed a clean solo line of the character and the voice is PINNED across the chain instead of re-performed from text. The bank carries voice too, but only from shot 2 on - this covers shot 1 as well. Trimmed to 15s; reference rows cost speed on every step.
sampler_overrideopt	STRING		Link a sampler NAME here (e.g. from H3 Studio Controls) to drive this widget from one master source. Overrides sampler_name when connected.
scheduler_overrideopt	STRING		Link a scheduler NAME here to single-source it. Overrides scheduler when connected.
self_anchor_voiceopt	BOOLEAN	false	AUTOMATIC voice identity: after shot 1 renders, its own audio becomes the reference (<Audio 1>) for every later shot - the voice the model actually performed is pinned, no file needed. Write shot 1 so the character speaks a clean solo line. An external voice_ref, if connected, takes priority.
reference_image_sizeopt	COMBO	match	Reference image sizing. 'match' scales each ref (down only, keeping aspect) to the generation's pixel area; 'max' uses the reference pipeline's 2048px short edge for best identity fidelity. Reference tokens ride through every sampling step, so 'max' can be several times slower.
preview_first_shotopt	BOOLEAN	false	Write shot 1 to output/video/H3_FIRSTSHOT/ the MOMENT it finishes decoding - minutes before the full chain completes - so a bad take can be cancelled early. The full path is printed to the console.
output_scaleopt	FLOAT	1.001–4	FINAL size multiplier, applied after decode. No upscale model: a lanczos resize, 1.0 is off. WITH a model: the model runs at its OWN fixed factor (usually 4x) and this brings the result to source x this value, so 2.0 on a 4x model gives 2x, not 8x. CAREFUL - 1.0 does NOT mean off once a model is wired; it means do-not-correct, so you get the full 4x. At 1344x768 that is 5376x3072: 94 MB a frame, 22 GB a shot, and every shot stays in system RAM until the master is joined. The console prints the projected size when the model loads - read it. Adds resolution, not detail. Works with every continuity mode; the bank still stores base-resolution clips.
sigmasopt	SIGMAS		Optional custom sigma schedule, replacing sampler/scheduler + steps entirely. Some turbo LoRAs ship a schedule they need in order to work at all. When this is connected the 'steps' and 'scheduler' widgets are IGNORED - the step count becomes len(sigmas)-1 - and the console says so. The two-pass upscale split is taken as a fraction of the supplied schedule.
save_every_shotopt	BOOLEAN	false	Write EVERY shot to output/video/H3_SHOTS/ the moment it decodes, in addition to the master. Insurance for long chains: everything that fails after the last shot - a mux OOM, a full disk, a cancelled tab - otherwise destroys the whole render at once. Shots are written BEFORE the seam trim, so consecutive files overlap by ~1s; the master is still the clean join. Costs one file write per shot.
upscale_model_nameopt	COMBO	(none)	Pick an upscale model by name instead of wiring a loader node. Synthesises detail rather than resizing, per shot, at the model's OWN fixed factor - usually 4x. Set output_scale to the size you actually want; leaving it at 1.0 lets the raw 4x through, which is rarely what you meant. The console prints the projected frame size and RAM cost when the model loads, before any sampling. Reads models/upscale_models/, the same folder ComfyUI's own Load Upscale Model reads.
master_normalizeopt	COMBO	luma+contrast	Deflicker the FINISHED chain: every frame driven to ONE global luma target, after the master exists. Per-shot correction cannot work here - it never reaches the raw-latent pin that carries the drift, and correcting shots against a rolling target leaves a step at every join (measured: all per-shot dials ON still gave +142% texture and +18% luma over three shots). This runs outside the feedback loop and lands every frame on the same number, so it cannot create a seam. Brightness only: texture drift is NOT fixable after the fact, because the only lever is blur and blur destroys real detail along with the invented kind.
pin_framesopt	COMBO	22	context_pin only: how many frames of the previous shot are pinned as raw latents at the head of the next one. This is the whole join. 22 is the shipped default; the longer settings hold the previous shot's composition further into the new one, which matters most at the FIRST join - shot 1 has nothing pinned behind it, so it is the only shot whose framing can disagree with the text. All four values are latent-aligned; arbitrary numbers are not.
pin_noiseopt	FLOAT	0.0500–0.1	context_pin only: mix this much seeded noise into the PINNED LATENT before it conditions the next shot. Same noised-clean-condition idea as join_anchor_noise, aimed at the thing that actually carries the drift here - measured, the texture ratchet under context_pin rides the raw latent pin, and neither join_anchor_noise (keyframes only) nor bank_ref_noise (bank images) touches it. Small, and scene-dependent: measured -1.8% per hop at 640x352 and -0.9% at 960x544 on a detail-heavy scene, against much larger gains on a scene that barely ratcheted at all. It cannot touch the dominant drift in a busy frame - master_normalize=luma+contrast is what does that. Above 0.10 it gets WORSE (0.20 measured 1.228 against a 1.211 control), which is why the range stops there. Set 0 to disable. The noised latent conditions the next shot but never reaches the final cut.
pin_renormopt	BOOLEAN	true	context_pin only: rescale each pinned latent so its standard deviation matches the FIRST pin's. The pin's own sigma climbs every hop (1.0325, 1.0368 against a 1.0220 shot-1 anchor), and that inflated pin is what conditions the next shot - so it compounds upstream of anything a master pass can reach. Measured against a matched control, same seed, 960x544 124f x4: texture growth over the chain +15.1% -> +11.5%, and framing correlation to shot 1 held 0.985 -> 0.996 by the last shot. The framing gain is the bigger one and cannot be a metric artifact - post passes do not move composition. One seed, one canvas. A scalar rescale moves no structure, so unlike a pixel correction it cannot blur detail.
reference_subjectsopt	STRING		How your reference pictures group into PEOPLE. Empty (default) = every reference picture is declared a photograph of the same one person. That is correct for a single character and WRONG for several - the model is told they are all the same individual and renders the average. Comma counts in picture order: '3,3' means pictures 1-3 are person A and 4-6 are person B; '2,2,2' for three people. Only <Subject 1> is described as speaking, because H3's voice conditioning is single-speaker.
reference_videoopt	IMAGE		NEW IN 2.2.4. Frames of an EXISTING clip, handed to the model as a video reference alongside the picture references. Read what this does before wiring it: H3 is told a video reference is 'an earlier moment of this same continuous scene' and to keep its framing, camera distance, room contents and colour temperature. It is SCENE and APPEARANCE conditioning. It is NOT motion transfer - there is no pose, depth or optical-flow path in H3, so it will not make your subject copy the movement in the clip. Feed it through H3ReferenceVideo, which trims and subsamples to the 2 fps the model actually reads; a raw 25-second clip is ~50 reference frames riding every sampling step.
reference_video_audioopt	AUDIO		Optional soundtrack for reference_video. H3 pairs video references with an audio reference, so if you leave this empty silence is generated to match. Supply the clip's real audio when you want its voice timbre referenced too.
low_ram_masteropt	BOOLEAN	false	Stream the master to disk instead of holding every decoded shot in host RAM until the join - peak RAM becomes ONE shot. Each shot is staged lossless the moment it decodes, levelled with the same master_normalize math (from stored statistics), and the finished file's path comes out of the new master_path output; master_frames carries a single placeholder frame. v1 streams the default config only: join_blend, join_fx and color_level=scene fall back to the RAM path with a printed reason. Needs ffmpeg on PATH.
audio_pin_framesopt	INT	00–240	context_pin only: frames of the previous shot's AUDIO to pin as reference, independent of the picture pin. 0 = same as pin_frames (22 = 0.9 s). Longer audio context costs conditioning rows but NO delivered frames - the head trim stays at pin_frames. 96 (4 s) is the audio-memory window the JoyEcho ancestor of this sampler carried between chunks; try it for continuous speech across joins. Experimental.
pin_noise_audioopt	BOOLEAN	false	EXPERIMENTAL: apply pin_noise to the AUDIO half of the pin as well, keeping the joint AV latent statistics consistent (2026-08-23 model-consult hypothesis for speech misalignment at joins). Field measurement says audio noising dulls the voice - that is why this defaults OFF. Flip it only for an A/B.
audio_tone_controlopt	BOOLEAN	false	The audio twin of chain flatten: chained audio drifts DULLER per hop (4-10 kHz fell 8-13%/hop even with the pinned slot, measured 2026-08-11). This EQ-matches every later shot's long-term spectrum to shot 1's - a constant linear filter per shot, clamped +/-9 dB, half-strength in the top band so it cannot manufacture hiss.
x0_texture_clampopt	FLOAT	0.0000–0.15	EXPERIMENTAL: during the LAST 30% of sampling steps, attenuate the high band of the model's x0-prediction by this fraction (video half only). Attacks the per-hop texture overshoot at the source, before it ever enters the pin. Too high reads as waxy shimmer; start at 0.02-0.05. 0 = off.

Outputs (7)

Name	Type	Description
master_frames	IMAGE	The joined master, seams trimmed.
master_audio	AUDIO	Master audio.
shots_rendered	INT	How many shots rendered.
video_latents	LATENT	Every shot's video latent EXACTLY as sampled, batched along dim 0 (one entry per shot), UNTRIMMED. Shots after the first open with head_frames of replayed material from the previous shot's tail - that is what makes the join seamless, and it is removed at decode, not in the latent. So this does NOT line up with master_frames until you trim it yourself. Latent temporal rows for F frames are 5*((F-5)//17)+2.
audio_latents	LATENT	The matching audio latent per shot, same batching, same caveat.
head_frames	INT	Frames of replayed head carried by every shot AFTER the first (0 for shot 1, and 0 for continuity modes that do not pin). Trim this many frames off the front of each later shot after you decode.
master_path	STRING	low_ram_master only: the finished master file on disk. Empty string when low_ram_master is off (use master_frames as always).