2.2.2 — the reference block was forcing glasses on everybody

One fix, reported by a user against 2.2.1. If you render with reference images, take this one.

Fixed: reference renders forced glasses onto the subject

Reported on Civitai: a reference image produced a character wearing thick black frames in every shot, and prompting to remove them changed nothing.

The pack injects a retention_analysis block alongside reference images so identity and voice stop drifting across a chain. That block was hardcoded to say the subject "retains the same face, skin, hair, glasses and wardrobe" - so every ref2va render was told to keep glasses whatever the prompt said. And it could not be argued with, because the sampler runs on a BasicGuider: cfg 1.0, no negative branch. There is nothing for a negative statement to subtract from, so "remove the glasses" only put the word into the conditioning a second time.

The block now says "the same face, skin and hair" and names no accessories. Eyewear, hats and jewellery belong to your prompt.

The rule this exposes, and it applies to every prompt you write for this graph: anything hardcoded into unconditional conditioning is permanent from your side. At cfg 1.0 you can add but never subtract, so phrase everything positively - "clear unobstructed eyes", never "no glasses". Negations do not work anywhere in this pack.

2.2.1 — three crashes, all of them after the expensive part

A bug-fix release. Four defects, three reported from real renders on real machines rather than found in review. If you have been unable to finish a long chain, or a high-resolution one died partway with the whole server going down with it, this is the release for you.

Fixed: two crashes that only appear on long or high-resolution chains

Both were DefaultCPUAllocator: not enough memory — system RAM, not VRAM — and both killed the job after it had already done the sampling.

master_normalize built the entire finished timeline four times over. A 12-shot 1088x1920 chain asked for a single 31.2 GB block and died on a 24 GB machine after 81 minutes of work. One of those four copies was pure waste: the code rebuilt a tensor byte-for-byte identical to one already in memory, purely to compute two numbers for a log line, then threw it away. Every statistic it needs is one-dimensional. It now measures per shot and releases each shot as it is consumed, so peak memory is one finished timeline plus one shot instead of four timelines. Output is bit-identical — verified against the old implementation across every colour mode and median width, on the pixels and on the log strings.
The upscaler was handed every frame at once. 243 frames at 1472x2560 is an 11.0 GB float32 allocation on top of the input. It now runs in chunks into a preallocated buffer, so peak is one chunk rather than the whole batch. Bit-identical, verified at two chunk sizes. Set H3_UPSCALE_CHUNK if you want to trade memory for a little speed; default 16.

Fixed: auto-reserve could clamp itself into a hard crash

On a 12-shot 1280x736 run, shot 1 loaded completely and rendered fine. Shot 2 carries the context pin and the reference rows, so its activation-pool requirement roughly doubles — the node correctly costed it at 18.2 GB. It then clamped the pool to 9.4 GB "so the weights still load completely", and the weights loaded partially anyway, 401 MB offloaded. Neither constraint was met and the render aborted inside a CUDA kernel, taking the whole ComfyUI process with it.

Two causes, both fixed. The keepout was 384 MB, which did not cover ComfyUI's own buffer plus the difference between what the driver reports free and what ComfyUI treats as usable — about 400 MB short. It is now 1 GB. And the "this is tight" warning compared the clamped reserve against a previous shot's measurement rather than what this shot had just asked for, so a 9.4 GB clamp looked fine against shot 1's 9.1 GB while the payload needed 18.2. It now compares against the real request and says plainly that this usually dies inside a kernel, and that raising the reserve cannot help because the memory is not there.

Worth knowing regardless of this fix: shot 1 succeeding tells you nothing about shot 2. Shot 1 has no pin and no reference rows. If shot 2 will not fit, lower frames_per_shot or the resolution, or load a smaller quantisation of the DiT.

Fixed: start_image on the memory sampler silently did nothing

Reported by a user. start_image on H3MultishotMemorySampler is an identity reference row, not a first frame — and on an fl2va checkpoint, which has no reference rows, it is built and then ignored entirely. Meanwhile the sibling node H3MultishotSampler has an input with the same name that really is an I2V first frame, which is where the perfectly reasonable expectation comes from. The combination produced no warning at all.

It now prints one, naming both ways to actually open shot 1 on a picture: use H3MultishotSampler, whose start_image is a true first frame, or set continuity=flf_chain here and feed keyframe_images (N+1 stills for N shots). Note that continuity=first_frame is not about your image either — it hands over the previous shot's last frame, and does nothing on shot 1.

Verified

Three chained shots rendered end to end on ComfyUI 0.32.0 with the new auto-reserve: every shot full load: True, which is exactly where the crashing run went partial, and the new warning fired on shot 2 with real numbers and did not stop the render. master_normalize ran on that same 685-frame chain. The chunked upscaler is proven bit-identical by test rather than by render — that graph had upscaling off, so I will not claim end-to-end coverage it did not get.

2.2.0 — everything since 2.1.2, in one post

Seven point releases folded into one version, because 2.1.2 is where most people still are. Five separate defects that stopped the workflow running, a measured campaign against drift over long chains that changed the shipped defaults, and a whole ComfyUI version this pack could not previously run on.

the Audio Spine produced static on ComfyUI 0.32.0
naming an mmproj file broke GGUF text encoders
The value 1 for reference_image_size is not available
the full workflow needed a node from a pack that is not in the zip
H3_Seamless_Chain_CORE could not be queued at all — the no-third-party-dependencies workflow, un-runnable since 2.1.3

New controls: pin_frames, pin_noise, pin_renorm, and master_normalize's luma+contrast mode. Changed defaults that change your output: memory_frames 2 → 0, and join_anchor_noise / handoff_release to 0 because they are inert under context_pin. New runtime: ComfyUI 0.32.0; 0.30.0 is unchanged and still supported.

Everything below is the original note for each version, newest first.

2.1.9 — CORE could not be queued, and bypassed nodes now say what they need

If you use H3_Seamless_Chain_CORE, it has been un-runnable since 2.1.3. Queueing it failed with:

value_smaller_than_min: Value 0.0 smaller than min of 1.0 - output_scale

Its sampler still carried the widget array it was saved with before 2.1.3 removed the four two_pass_upscale dials. Nineteen saved values against the class's fourteen live widgets, so the frontend put false into output_scale (minimum 1.0) and 1.5 into save_every_shot, and the server rejected the whole prompt. The workflow advertised as the one with no third-party dependencies — the safest thing for a new user to open — was the one that could not run.

It survived because CORE had never been rendered end to end. Our own release checklist said "one CORE render before posting" and that line rode along, unticked, through six releases.

The array is now generated from a name→value map resolved against the server's own schema, with the map stored on the node so it can be re-applied by name. CORE has now rendered: three chained shots at 960x544, 370 frames, picture and audio, holding framing, wardrobe and lighting across both joins. All three bundled workflows are submit-tested against a running server now, not read.

Bypassed nodes now say what they need

Four nodes ship bypassed because their packs are not in this zip, and a node whose class is missing fails the entire prompt. That is the right default — but a bypassed node sitting on the canvas invites you to un-bypass it, and doing that without the pack installed breaks the workflow with no explanation.

It now says so in four places: the group title, each node's own title (which names its pack), a note above the cluster, and the VRAM / SPEED note, which carries the full table with repository URLs.

attention patch, chunk FFN — ComfyUI-sol-attn
block cache — comfyui-minimax-h3-blockcache-T8
SCRIPT PREVIEW — ComfyUI-Custom-Scripts. A leaf: leaving it bypassed costs you the on-canvas preview and nothing else.

The part people lose is that it takes two steps, in order: install the pack, Ctrl+B the node to un-bypass it, then turn its switch on. Un-bypassing alone does nothing; flipping the switch while the node is still bypassed does nothing either.

Why not bundle those packs? The prompt-writer pack is bundled because it is modified and pinned to this workflow. These three are not — shipping a second copy of Custom-Scripts, sol-attn or blockcache-T8 inside the zip would shadow whatever you already have installed and freeze it at whatever version we happened to vendor. Naming them and linking them is the honest version.

2.1.8 — the last node that came from outside the zip

SCRIPT PREVIEW is ShowText from ComfyUI-Custom-Scripts, and it shipped active. A node whose class is missing serialises as class_type: null and the server rejects the whole prompt — so without that pack installed the workflow could not be queued, for the sake of an on-canvas text box. The install notes said to delete the node, which only helps someone who reads them before pressing Queue.

It now ships bypassed, the same remedy the three accelerator nodes got in 2.1.4. Install Custom-Scripts and Ctrl+B the node if you want the preview; the writer feeds the sampler either way.

Found by mapping every node type in all three bundled workflows to its owning Python module against a running server — now part of the release routine rather than something I do after someone tells me. The other two workflows were already clean, and everything else in the full one is core ComfyUI, this pack, the writer pack that is in the zip, or one of the three bypassed accelerators.

Verified end to end on ComfyUI 0.32.0

Not by reading the graph — by loading the shipped file in a browser on a 0.32.0 install and pressing Queue. Twice: once through the prompt-writer lane exactly as shipped, once with a hand-written script.

loads with no missing node types, and every sampler widget lands on its own name — the shift behind The value 1 for reference_image_size is not available cannot reproduce
no null class_type: the bypassed preview and the three bypassed accelerators all drop cleanly
three chained shots at 960x544, 124 frames each, decoding to 328 frames — exactly 372 minus the two 22-frame context_pin head trims
a reviewer given the clip cold, with no idea how it was made, read it as one continuous static take, found no cut or jump anywhere, transcribed every line, called the lip-sync matching, and found no shift in framing, colour, brightness or wardrobe and no hiss, dropout or click in the audio

0.30.0 remains the version everything else here was measured on. Both are tested before release now.

2.1.7 — three things that stopped it running

If you installed from Civitai, you have a package with three separate blockers in it. All three are fixed here. Two of them only ever appear on ComfyUI 0.32.0, which is why they survived several releases — my test rig was 0.30.0 and could not reproduce them. Everything below is now verified on 0.32.0 as well.

Fixed: the Audio Spine produced static on ComfyUI 0.32.0

guide_audio came out as hiss, while the same file through voice_ref or the native node was perfect. Reported against 2.1.2 and still present through 2.1.6.

0.32.0 introduced ModelSamplingAV, which carries the audio half of the packed audio+video latent scaled onto the video schedule — multiplied by shift / audio_shift going in (12/3 = 4 for H3) and divided back coming out. Everything this pack injects into the sampler's latent is in the stream's native domain, so on 0.32 it landed four times too small and decoded as broadband noise. On 0.30.0 there is no such scaling, so the identical code was correct there.

Three injection sites had it, not one: the encoded spine, the previous shot's audio tail (which audio_lock and the latent handoff use), and the encoded room tone for the onset guard. All three now scale at the point of use, reading the factor off the live model object so a changed sigma shift stays correct, and leaving 0.30.0 byte-identical.

Verified by render on 0.32.0 with a real 44.1 kHz stereo voice track: loudness-envelope correlation against the guide +0.965, speech-band energy 34.2% against the guide's 39.3%, and a blind listener transcribing the guide's own words as "clean, no background hiss". The same render was hiss before the fix.

Fixed: naming an mmproj file broke GGUF text encoders

mat1 and mat2 shapes cannot be multiplied (3680x1152 and 3456x1152)

— thrown at the handoff into shot 2, on GGUF encoders only, and unaffected by resolution or by switching every image input off.

Setting mmproj_name explicitly took a different code path than (auto) and skipped the vision key-renaming step, so the vision tower loaded under raw llama.cpp names that nothing downstream reads. Same file, two loaders, 19 key names in common out of 351. The bitter part is that mmproj_name is documented as the reliable escape hatch for when filename pairing fails — and it was the broken path.

It now runs the same post-processing as (auto), using ComfyUI-GGUF's own key map rather than a copy, so it follows their changes. Verified: an explicitly named file now produces a state dict identical to (auto). A new guard also fails by name if a chosen mmproj yields no vision tensors, rather than dying in a matmul twenty minutes into a render.

Fixed: The value 1 for reference_image_size is not available

The shipped workflow's saved widget values were written for a layout that did not present sampler_override and scheduler_override as widgets. The current schema does, so everything from index 27 read two slots early and output_scale's 1.0 landed in reference_image_size. This affected 2.1.3, 2.1.4 and 2.1.5. The array is now generated from a name→value map resolved against the server's own schema, with that map stored in the node so it can be re-applied by name if anything shifts again.

Also: memory_frames now defaults to 0

The memory bank's recent slots hand each shot's accreted output forward as reference images, on top of the latent pin, so invented detail compounds. Ten shots at 960x544, moving 2 → 0:

                              memory_frames 2    memory_frames 0
texture per hop                    1.055              1.022
chroma per hop                     1.086              1.039
framing correlation at shot 10     0.976              0.995
drift acceleration            4.2 -> 6.7%/hop    2.3 -> 2.0%/hop

The last row matters most. At 2 the drift accelerates, which is what a runaway feedback loop looks like; at 0 it holds flat.

The obvious worry was motion continuity, since carrying it is what the recency slots are for. Tested on a scene with continuous large-amplitude movement: anchor-only retained motion slightly better (−5.9% against −6.5% over four shots) with better framing (0.983 against 0.971). The cost does not exist. If a busy scene ever does lose continuity between shots, raise it to 1.

join_anchor_noise and handoff_release now ship at 0. Both are inert under context_pin — one noises keyframes the mode never creates, the other belongs to latent_handoff — and non-zero values read as tuned settings while doing nothing at all.

2.1.6 — chained shots stop getting brighter-edged

master_normalize matched every frame's mean to one global target, which is why a chain shows no brightness step. It was also masking a second drift it never touched.

Measured on a three-shot chain at 960x544: the mean held flat, 27.31 → 27.43, while the distribution stretched — p25 fell 6 → 2 and p95 rose 85 → 96. Contrast climbing every hop, re-centred each time, and handed to the next shot as a higher-contrast start.

New: master_normalize=luma+contrast, now the default, matches the spread as well as the mean. Rescaling amplitude about each frame's own mean is an affine remap — it moves no edges, so it is not the blur this pack has always ruled out for texture drift.

Texture growth per hop, from a log fit across all shots. 1.000 is no accretion at all:

where                          luma    luma+contrast   contrast spread
960x544, in-render, 124f x4   1.126       1.047        11.2% -> 0.3%
960x544, 243f x3              1.199       1.064        12.2% -> 0.4%
640x352, 243f x3              1.130       1.055         7.0% -> 0.2%

The baseline ratchet scales with canvas (1.130 at 640x352, 1.199 at 960x544); after normalising it stops caring (1.055 against 1.064). There is nothing in it to tune per resolution — it works on decoded frames with per-frame statistics against one global target.

The target anchors to shot 1, not the timeline median. Contrast only ratchets upward, so shot 1 is the one frame-set with nothing accreted onto it. A median target pulls shot 1 up to meet the drift (+11.7% texture, for no benefit); anchoring to shot 1 leaves it untouched (+0.1%) and only ever pulls later shots down.

1:1 crops of the last shot show no loss of real detail — lamp vent slots, hinge rivets, hair strands and knit weave all survive. What leaves is the invented crispness.

What is left, honestly

About 1.05 per hop. That residual is spatial accretion and this pass cannot reach it: master_normalize runs on the finished master, outside the feedback loop, so it cleans what you see while the next shot is still handed the inflated pin. Four shots is slight. Ten shots is roughly +50%. If you are chaining long, expect it.

Correction to 2.1.5

2.1.5 said pin_noise=0.05 fixed this. It does not. That was measured on two seeds of a single scene at 640x352 — a scene with a nearly black background that barely ratcheted to begin with — and the control that would have caught it, pin_noise=0.00 at the reporting user's own resolution, had never been run. With it run:

canvas      0.00    0.05    change
640x352    1.131   1.111    -1.8%
960x544    1.211   1.201    -0.9%

Both on a detail-heavy scene. The dial is small and scene-dependent, it cannot touch the dominant drift in a busy frame, and above 0.10 it gets worse (0.20 measured 1.228 against a 1.211 control). Its range now stops at 0.10 and its tooltip says so. It stays in the pack because it costs nothing and does help where the ratchet is already small — it is not the fix.

Not tested

Portrait canvases. Everything above is landscape, 640x352 and 960x544. The mechanism is resolution-independent by construction and the two landscape sizes agree, but 768x1344 and 736x1280 have not been measured and are not claimed.

If you measure this yourself

Texture comparisons only mean anything while the framing holds. If the model cuts to a different setup, texture reflects content and the number is meaningless — one portrait run here scored a flattering 0.878 per hop purely because shot 3 cut to a close-up of a film reel. Correlate each shot's mean frame against shot 1's before trusting any of it; a held framing sits above 0.95.

That cut is worth a writing rule of its own: do not name a nearby object in a shot's closing beat. “She glances down at the reel” reads as a request for a shot of the reel. Keep closing beats on the speaker's own body.

2.1.5 — widget slots, and two dials

Fixed: the shipped workflow would not queue. The value 1 for reference_image_size is not available. The saved widget array was written for a layout that did not present sampler_override and scheduler_override as widgets, so everything from index 27 read two slots early — output_scale's 1.0 landed in reference_image_size, a combo of match/max. This affected 2.1.3, 2.1.4 and 2.1.5. The array is now generated from a name→value map resolved against /object_info, and the same map is stored in the node's properties so the pack's own JS can re-apply values by name if a future schema change shifts anything again.

New: pin_frames (default 22) exposes the pin length, previously hard-coded. Only the Motion-Context node's latent-aligned values are offered: 22 / 5 / 39 / 56. Longer pins cut the ratchet a little more, but the head trim scales with them, and a blind review of the 39 arm caught a jump at the third join with a line clipped from “So it stays in the drawer” to “…is in the drawer”. Leave it at 22.

2.1.4 — two user-reported breakages

Fixed: the main workflow would not queue without two third-party packs

H3_Seamless_Chain_v2.json carries three optional accelerator nodes — sol_attn and chunk_ffn from ComfyUI-sol-attn, and block_cache from comfyui-minimax-h3-blockcache-T8. They were saved active, so on an install without those packs the graph could not be queued at all:

missing_node_type: Node 'attention patch (gated)' has no class_type

— even though the shipped recipe keeps all three gates off and never touches them. They now ship bypassed (purple). A bypassed node is dropped from the prompt entirely and the model passes straight through, so the workflow queues with neither pack installed. To actually use one: install its pack, select the node, Ctrl+B to un-bypass, then turn its gate switch on. Both steps.

Fixed: Value 4 bigger than max of 3: memory_frames on a graph you never edited

Mine, not yours. v1.2 inserted a new seed_per_shot control into the middle of the sampler's input list. ComfyUI stores widget values as a positional array, so in any workflow saved on v1.0 or v1.1 every dial after that point shifted by one — your old anchor_frames was arriving as memory_frames. The error named a dial you had nothing to do with.

Fixed in both directions. Pre-v1.2 workflows are detected as they load and repaired (save the workflow to make it permanent), and the sampler now also stores its values by name, so no future change to the input list can shift anything again. If a bad value still reaches the queue, the validation error explains the shift instead of pointing at the wrong dial.

2.1.3 — the upscale, honestly

two_pass_upscale is removed. It interpolated the raw latent between passes. H3's latent is not a spatially smooth representation, so interpolated values landed off-manifold and the second pass, running at low sigma, had no room to pull them back. Every configuration tested came back as colour-noise mush against clean single-pass controls — including at 14 steps with beta57, the recipe the settings file itself called render-verified, and including shot 1, which carries no pin at all. It was never a context_pin incompatibility. It did not work in any mode, and the guard that used to point at context_pin is gone with it.

Two replacements, both applied after decode, where nothing can leave the latent manifold, and both per shot, so a long chain never holds a full upscaled master in memory at once — which is exactly the failure that cost one user a three-hour render.

output_scale — a lanczos resize of each shot's finished frames. It adds resolution, not detail, and the measurement is worth knowing before you choose: rendering at 448x256 with output_scale 1.5 reached 672x384 in 45.5 s against 80.9 s rendering 672x384 natively — 1.78x faster, and visibly softer. Concrete pore texture that survives a native render washes out. Use it when the clock matters or the whole-chain upscale will not fit; render native when texture matters.
upscale_model — optional UPSCALE_MODEL input (ComfyUI's own Load Upscale Model — ESRGAN and friends) to synthesise detail instead of resizing, per shot, at the model's own factor. Implemented but not render-verified: no upscale model was installed on the test rig, and I would rather say so than imply otherwise.

Both run after the memory bank has taken its base-resolution reference clip, so conditioning and VRAM are identical to an un-upscaled run, and the latents the node returns stay base-resolution.

New: the sampler returns its latents

The memory sampler now outputs video_latents, audio_latents and head_frames — every shot exactly as sampled, batched along dim 0, untrimmed. Shots after the first open with head_frames of replayed material from the previous tail, and that is only removed at decode, so these do not line up with the master until you trim it yourself. They are deliberately raw: the pin material cannot be recovered later, and trimming on your behalf would throw it away. Requested in issue #12, and the requester picked exactly this shape.

Verified rather than asserted: 124-frame shots return 37 latent rows, matching 5*((F-5)//17)+2, and 124 + 124 − 22 is exactly the 226-frame master.

If you saved your own copy of a v2.1.2 graph, reload the shipped workflow. Removing four widgets shifts the saved widget order on that node.

2.1.2 — reference images, per-shot saves, and six repairs

One new capability and a set of fixes. The chaining itself is untouched: nothing here changes how shots join, so a graph that renders well today renders identically after updating.

New: reference images actually have a way in

The sampler's reference_images input has existed since 2.0, and SETTINGS.md documented it — as unwired, because nothing in the shipped workflow was connected to it. There was no way to use it without building the lane yourself.

The full workflow now carries a REFERENCE lane in the anchors column: two image loaders → ImageBatch → a gate → the sampler. It ships with the gate off, so nothing changes until you turn it on. Point the loaders at portraits of your character, flip the gate, and they ride into every shot as <Picture 1>, <Picture 2>. Chain another ImageBatch for a third and fourth. Needs a ref2va checkpoint.

Unlike the identity anchor above it, these are not a first frame — they do not constrain shot 1's composition, they only carry who the person is. That also makes them the thing that covers shot 1, where the memory bank is still empty and has nothing to carry identity from yet.

New: every shot can be saved as it renders

A chain only became a file at the very end, when the master was muxed. So anything that failed after the last shot destroyed the entire run — and one report was three hours lost to an out-of-memory error at the mux, after every shot had already rendered successfully. The work existed and was thrown away, which is the worst version of that bug rather than the mildest.

save_every_shot (on both samplers) writes each shot to output/video/H3_SHOTS/ the moment it decodes, alongside the master. If the mux dies you have every shot on disk and a joining job instead of a lost day. Files are written before the seam trim, so consecutive shots overlap by about a second — the master is still the clean join, these are the safety copy. Requested in issue #13.

New: custom sigma schedules

The samplers built the schedule themselves from steps + scheduler, with no way to supply your own — so a turbo LoRA that ships the curve it needs could not use it, and ran wrong rather than refusing. Both samplers now take an optional SIGMAS input. Connect one and it replaces the schedule entirely, steps rebinds to len(sigmas)-1 so the two-pass upscale split is taken as a fraction of your curve, and the console states that the steps and scheduler widgets are being ignored instead of silently overriding you. Link-only input, so saved graphs load unchanged. Issue #14.

Fixed: --- separators were ignored in passthrough mode

example_script.txt ships --- separated and every document tells you to write scripts that way, but the writer's passthrough path returned the whole file as ONE shot — which the sampler then repeated to fill shot_count. Pasting a finished four-shot script rendered the entire text as shot 1, four times over. It now splits on exactly the same rule the sampler uses. A single paragraph is still one shot, so .txt batches are unaffected.

Fixed: a stale prompt-set filename killed the whole queue

ComfyUI validates every combo value in a graph before it runs anything. If RiftPromptSource's saved source_file no longer existed — a renamed folder, a workflow shared from another machine, or the prompt lane simply switched to manual — the queue failed with Value not in list and nothing ran, including the lanes that were perfectly fine. Switching to manual did not disable it, because validation happens before the switch is ever consulted.

The node now declares VALIDATE_INPUTS, so the filename is resolved only if the node actually executes. Manual mode now genuinely disables it. If it does run and the file is missing, the error names the file.

Fixed: the writer

These three only matter if you let the LLM write your shots. If you paste your own shot list, they change nothing for you.

Every story came out 15 shots. The system prompt ordered exactly 15 whenever the brief didn't ask for a count, so the model never got to decide. It now counts the story's beats and lands where the story lands — measured 4–7 on ordinary briefs, and about 7 when the brief gives no length signal at all. Seven chained shots is roughly 65 seconds at 243 frames, which clears the 1-minute mark platforms pay on. Ask for a specific count and you still get exactly that count.
The mode dropdown did nothing. The shipped workflow's system_prompt box held a frozen copy of the long-story prompt, and a filled box overrides the per-mode prompt file — so every mode ran the long prompt no matter what the dropdown said, and prompt updates in the pack never reached anyone using the shipped graph. The box now ships empty and the dropdown works. If you saved your own copy of the v2.0 or v2.1 workflow, clear that box by hand — your saved graph still carries the old frozen prompt, and this update cannot reach it. Related: short_story is now 1–3 shots instead of always exactly 1.
A messy answer from the model no longer kills the render. Three separate real failures, all seen on local models: a markdown fence sharing a line with the payload (```json {"prompts":…) was destroyed by the fence stripper; a reply truncated at the token ceiling left no balanced object to recover; and parsing sat outside the retry loop, so one malformed answer ended a run after a 100–240 s call that a re-ask usually fixes. The order is now clean → parse → retry ×3 → salvage the shots that completed → fail, and the final error names passthrough mode as the escape hatch.

Verification, plainly. The reference lane, the per-shot saves and the --- fix are all confirmed by a live chained render, not by inspection: the console logs 2 reference image(s) ride in every shot as <Picture 1..2>, PASSTHROUGH: 4 shot(s) (it would have said 1 before the fix), and shot 1/4 saved through shot 4/4 saved, with four files on disk. The three JSON failures are unit-tested against the actual captured payloads that caused them; the live re-run afterwards parsed on the first attempt, so that proved no regression rather than proving the salvage path fires in a real queue. Shot counts were measured over two runs of five briefs on a local 8B: 4–7 shots, ±1 between runs — before the change all five returned 15. The custom sigmas input has not been exercised by a render — it needs a schedule source wired in, and I would rather say so than imply it.

Fixed: the Audio Spine produced static with real audio files

User-reported on ref2va: a voice track through guide_audio came out as noise, while the same file worked through the native reference node. The spine encoded the track at whatever sample rate the file arrived in; the audio VAE expects 32 kHz, and the native node resamples where the spine path did not. Nearly every real voice or music file is 44.1/48 kHz, so the spine's latent was garbage - and the spine locks every sampling step to it, which is why the output was noise rather than merely degraded.

Fixed: the spine resamples to the VAE's rate and upmixes mono to stereo, with a console line saying so. Measured, not asserted: guide-to-output correlation went from 0.06 to 0.97 on a 48 kHz voice track - identical to a native-rate control - and the same at 44.1 kHz mono. While in there, the tooltip's claim that the spine is latent_handoff-only was simply wrong: it works with every continuity mode and is render-verified on context_pin. If you want the music-video path, this is it: wire the song into guide_audio and every shot's audio locks to its slice of the track.

Measured: audio dulls over long chains — and the counters

The docs have long carried a claim (from upstream notes) that audio gets duller the longer the chain runs. Measured, five arms of 8-shot chains, one factor at a time: it is real, and it is the audio twin of the seam sharpening ratchet running the other way. Each shot's audio is regenerated conditioned on the previous shot's; regeneration regresses toward the spectral mean; the output becomes the next shot's conditioning.

With bank_pinned=0 — pure recency conditioning — the voice band collapses: 84–92% of 4–10 kHz energy gone by shot 8. There is no true "bank off": zeroing both bank widgets leaves exactly one recency slot, which is the worst configuration, and the node now warns about it on chains past 4 shots.
With the default pinned slot the drift is seed-dependent — 8% to 50% over 8 shots across the seeds tested. The pinned slot slows the loop; it does not break it.
audio_tone_control = flatten (new, default off): EQ-matches every shot's long-term spectral envelope to shot 1's before the weld — constant per-shot gains, clamped ±9 dB, half-strength in the top band so it cannot manufacture hiss. Paired A/B on the worst seed measured: HF loss halved (−49.5% → −23.7%), rolloff drift cut to a third. It reduces the drift, it does not eliminate it — the context_pin replay carries raw latents the EQ cannot reach — and it ships off until ears, not spectra, have judged it. (First ear test: marginally preferred, "but close.")
The best-audio recipe needs no new code: continuity=cut + bank_pinned=1, memory_frames=0 gives every shot exactly one audio reference - shot 1, forever. Measured flat (−5% over 8 shots, both seeds tested) where the seamless default lost 8–50%. The joins are cuts rather than a continuous take; identity and scene still hold. Pick by grammar: interviews and episodic content can have the flat audio, one-take content pays the pin's audio tax and levels it with flatten.
Picture darkening over chains: same mechanism, and the dial already existed. A user-grade chain measured −1.5 luma per shot, monotonic. color_level=mvgd — shipped since 2.1 but never verified — holds an 8-shot chain to −1.0 total where uncorrected loses −10.5, at both seeds tested. On chains past ~5 shots, turn on all three drift dials: chain_gain_control=flatten (texture up), audio_tone_control=flatten (sound down), color_level=mvgd (brightness down). One mechanism, three lanes — the model regenerating from its own output.

Correction, 2026-08-12. An earlier build of these notes recommended turning on three per-shot drift dials. A render with all of them on came out +142% texture and +18% brighter over three shots, with a visible brightness step at each join. Per-shot correction cannot work: under context_pin the drift is carried by the raw latent pin, stored before decode, while every one of those dials operates on decoded frames. They fix what you see, not what feeds forward. audio_tone_control is removed; color_level=mvgd is deprecated (29% warmth step at every join, per its own source comment). Use color_level=scene and the new master_normalize=luma — both run on the finished master, outside the loop, driving every frame to one global target, so neither can create a seam. Texture drift has no honest after-the-fact fix: the only lever is blur, and blur removes real detail along with the invented kind.

Install

Unzip. Copy both node folders into ComfyUI/custom_nodes/:

ComfyUI-H3-Multishot/            the sampler and helper nodes
ComfyUI_JoyAI_Echo_GGUF_Nodes/   the LLM prompt writer (full workflow only)

Restart ComfyUI. ComfyUI v0.30.0 or newer is required — that is the release with native MiniMax-H3 support.
Load a workflow from workflows/ through the workflow menu.

The full workflow needs five packs (two are in this zip): the writer pack below, ComfyUI-H3-Motion-Context for the context_pin default, RES4LYF for the beta57 scheduler, and ComfyUI-sol-attn + comfyui-minimax-h3-blockcache-T8 for the VRAM panel; ComfyUI-Custom-Scripts adds the script preview. ComfyUI validates every node class before it will queue, so a missing pack stops the whole workflow rather than just its own feature — INSTALL.md lists how to remove each one instead if you would rather not install it. CORE needs none of them, and that is tested on a clean install.

On the Motion-Context fork. There is an active fork, ethanfel/ComfyUI-MiniMaxH3-Contex-Loop, well ahead of upstream with a disk-backed chain/loop system. It is a complement, not a replacement: by design it leaves the MiniMaxH3MotionContext node id to NikoDemon80's pack. Install the fork instead of upstream and context_pin still fails, because the node it calls is not there. Install both - they are built to coexist, and this pack works with either one's runtime patches. As of 2.1.1 the error message says so directly when it detects the fork.

The writer pack is RealRebelAI's (github.com/RealRebelAI/ComfyUI_JoyAI_Echo_GGUF_Nodes), modified so the workflow's join rules actually reach the model; NOTICE_RIFT_MODIFICATIONS.md inside it lists every change. If you already have that pack, replace it with this copy. The CORE workflow does not need it at all.

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  (full)  /  beta  (CORE)
steps         14              fps         24
frames/shot   362  (~15.1s, the trained maximum)
resolution    1280x736 landscape   or   768x1344 vertical

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, with image quality, skin texture, artifacts and audio judged equal — so the full workflow ships it and lists RES4LYF as required, while CORE ships beta and keeps its zero-third-party-pack promise. One widget either way.

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 <Picture 1..N>. 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.

Fixed in 2.1.1

context_pin died when ComfyUI-H3-Motion-Context was installed. Both packs patched the same ComfyUI method and Motion-Context refuses to stack on an unrecognised wrapper, so its payload patch failed and the chain errored. This pack's wrapper now does everything theirs does and declares their compatibility marker, so whichever loads first owns the site and the other stands down. Load order no longer matters. Verified with a live context_pin render.
seamless_tail crashed mid-chain with Motion-Context installed ("only first/last keyframe anchors are supported") - after your first shot had already rendered. It needs interior keyframe anchors, which conflict with that pack; it now stops before any sampling with the alternatives named: use context_pin, or first_frame, or remove that pack.
seamless often reads as a cut - now labeled. It is a legacy latent-only soft pin kept for comparison; the model satisfies it loosely. For a real join use context_pin or first_frame. The tooltip and settings reference now say so plainly.

Why these never showed in testing here: an install-layout difference disabled the conflict detection on the dev machine. That detection is fixed, and release testing now runs on a packaged clean install so the class cannot slip through again.

Everything you need is in the zip. One download: both node folders, all three workflows, and the full documentation. Nothing else to fetch, nobody to ask.

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).

Files with SHA256: 3693e2135354831dba58bbcb4e4488d7837989e66d2874c1c6ba8071380e256f