Skip to content

Releases

H3 Studio 9.3.2 — Chapter sequences and render pause

Choose a tag to compare

@erdinoral erdinoral released this 18 Oct 12:08

H3 Studio 9.3.2 — Chapter sequences and cooperative render pause

Director chapter sequences

With Continuum enabled, Sequence grouping → One video per chapter processes the chapter's scenes as ordered native chunks. Six five-second scenes produce one 30-second sequence, with each scene's own prompt and reference assignment. Only the first starting image is applied; intermediate New-cut starting images remain saved but are not applied in this mode. Split at new cuts retains the previous behavior and remains the default for existing projects.

The native limit remains 16 chunks and nine unique reference images per sequence. A longer chapter splits into ordered sequences; the next sequence uses the previous output's last frame, which is not saved native motion/audio context. There is no promise of seamless continuity between separate jobs. Existing queued jobs are unchanged.

Video details and continuation choices show the total native sequence duration while running, rather than only its chunk duration. Scene cards show the actual configured seconds instead of a fixed one-minute label. Spectrum blend buttons and inputs in the mirrored Scene settings now forward to the existing settings handlers.

Render pause (#19)

Pause render, next to Stop, waits at a sampling progress checkpoint and resumes the same prompt execution. It does not interrupt, submit a replacement render, modify tensors, wrap the model, or change native saved-prefix signatures. Pause after current render remains a separate queue control.

An in-progress step must finish before the pause is acknowledged. Model/latent memory remains allocated; this does not free VRAM or suspend a model-loading stage. Stop remains effective while paused. Studio renews a 30-second lease; if the controller disconnects, the engine automatically resumes after lease expiry. It is not a durable pause across an application shutdown. The boot shim installs the bundled control integration on restart, without model downloads or edits to ComfyUI core files.

Production diagnostics (#28)

Open a video's Prompt → Diagnostic report to download a small JSON report with Studio status, cancellation source/time, model names, effective Spectrum settings and available engine-completion evidence. It excludes prompt text, reference images, local paths, API keys and raw logs, and does not upload anything. Missing history is reported as unavailable, not as successful completion. Exporting a report does not change job status or restart a render.

The 9.3.1 cancellation/completion race fixes remain in place. The original no-input CANCELLED report and visual drift still require evidence from the affected installation; this release does not claim that a local successful render establishes their cause.

Validation

239 Python tests and 46 JavaScript tests pass, with syntax and whitespace checks. A real Director native two-section draft at 864×480 completed as one eight-second, 192-frame video after pausing for six seconds with an unchanged sampler step and resuming the same prompt ID. The status finished as done. This is functional validation, not cinematic-quality or speed acceptance.

Hybrid selection (#21) was verified through the live API: both H3 model slots changed together and encoder/VAE choices stayed intact. The original model selections were restored. No model download or feature-film render was performed for this check.

Update through Pinokio, stop/start Studio when your queue is empty, then reload the browser. Existing projects and outputs are retained.

H3 Studio 9.3.1 — Director continuity and production reliability

Choose a tag to compare

@erdinoral erdinoral released this 18 Oct 10:59

H3 Studio 9.3.1 — Director continuity and production reliability

Director scene references could point to the wrong image when a take used different character subsets or a first-frame guide. Reference labels now resolve to fixed Continuum V39 slots, preserving each scene's own reference order while the native sampler handles keyframe offsets.

Director and generation

  • Preserve imported scene order, distinct per-section prompts and New/Continue choices; AI Director plans a progressing story rather than repeating an introduction.
  • Keep first-frame guides on New takes. Independent Continue takes require their own preceding Director scene, rather than the latest unrelated video.
  • New Continuum reference geometry defaults to Match Output. Max Identity remains an explicit Director API choice; existing saved selections are retained.
  • Production details distinguish saved native context, chunk count and last-frame handoff. Cross-job last-frame handoff does not transfer native motion/audio history.
  • Saved native-prefix continuation retains its original keyframe contract, including prompt-only runs without a first frame; it does not introduce the parent video's last frame into that saved run.
  • Retrying a New/image-guided job preserves its selected starting image. Independent continuation retries refresh their parent last-frame handoff; saved-prefix retries retain their original guide.
  • Omit identical Action/Beats repetitions without removing distinct chronological beats.
  • Spectrum presets and manual video blend preserve unrelated settings. Audio blend remains zero. Spectrum runs still cannot safely reuse persisted prefixes in this integration.
  • Record the Spectrum settings actually submitted to the engine in each job and its Gallery entry, independently of later settings changes.

Issue-related fixes

  • Stop cannot accidentally target the following job during an asynchronous interrupt request. Completion during Stop is preserved; explicit cancellations record source/time.
  • Legacy CANCELLED results can be recovered on startup only when Comfy confirms successful completion with a saved video. An MP4 alone is insufficient evidence. This does not establish the original cause of issue #28 or resolve visual drift.
  • Hybrid and model catalogs recognize subfolders with correct relative loader paths. Known full-size Hybrid weights and alternate filenames are supported; incomplete files stay unavailable. Both H3 model slots change together, retaining encoder/VAE selections.
  • Pause remains after the current render, not mid-render suspension. Issue #19 remains open.

Validation and limitations

All 225 Python tests and all 39 JavaScript tests passed, with syntax and whitespace checks. Passing code tests is not proof of cinematic continuity, speech quality, or a general speed increase. No personal reference images or generated film files are distributed.

Local RTX 5060 Ti 16 GB smoke tests completed at 864×480, 8 steps, native audio and no LoRAs: two native 2×4-second runs and one two-job last-frame chain. Each assembled result contains exactly 192 frames / 8 seconds and 32 kHz audio. These are draft functional checks, not final-quality acceptance. Sampled frames retained the main appearance, but the fixed-camera instruction was not fully respected; Spectrum 0.60 examples showed translucent/double-image artifacts. An initial exact-prompt Spectrum-OFF control was interrupted when Pinokio, Studio and ComfyUI became unavailable; that attempt was preserved. A fresh control subsequently completed after restart, as detailed below. The user's feature-film production was not resumed or published. Audio presence was checked; listening quality was not assessed.

After the user reopened Pinokio, a fresh exact-prompt OFF control completed: the actual prompt/chunk prompts, seed, steps, dimensions, references, guide and reference sizing matched the Spectrum 0.60 case. Both outputs contain 192 frames / 8 seconds and 32 kHz audio. Recorded wall time was 344.45 s OFF versus 272.86 s ON; OFF followed a restart with different loading/cache conditions, so this single draft pair is not an isolated Spectrum speedup benchmark or a 1088p prediction. The sampled OFF frames did not show the conspicuous translucent double-body trails seen in ON, but walking/fixed-camera compliance still failed full acceptance. This is a limited observation, not a universal causal conclusion. The interrupted attempt remains preserved. Original active film and Spectrum 0.60 settings were restored, the queue is idle and feature production remains paused. Audio listening and an equivalent 1088p benchmark remain outstanding.

Update through Pinokio, stop/start Studio, then reload the browser. Existing projects, media and saved production settings are preserved. No new model download is required for these fixes.

H3 Studio 9.3.0

Choose a tag to compare

@erdinoral erdinoral released this 13 Oct 04:36

H3 Studio 9.3.0

Video workflows

  • Experimental native Continuum generation and continuation in Scene; optional Continuum backend in Director. Existing Multishot remains the Director default because framing and identity can still drift.
  • H3 Refine in the player offers only resolutions higher than the source and saves a separate Gallery video. Enable capture before generating the source; existing MP4 files without captured state are not supported.
  • Refine uses the H3-trained 3D latent upscaler and conservative Flow-Aligned resampling. Optional node versions and model weights are pinned by Film tools. See workflow details.

Interface and sponsors

  • Detailed Turkish and English Scene/Director guide, simpler scene editing, and clearer save status.
  • Diamond204a is our first sponsor, contributing $25 and useful development feedback.
  • A red scrolling banner above the player alternates sponsor thanks and Buy me a coffee invitations every ten seconds, with pauses between messages.

Validation and updating

183 Python tests and 32 JavaScript tests passed. Local 480p draft Continuum and 480p-to-608p learned Refine renders completed. Refine preserved the original file; inspected frames retained broad framing and lighting without the earlier obvious double-edge artifacts. Facial expressions can change; audio was not assessed.

Run Update in Pinokio, then stop/start Studio and reload the page. Film tools installs the optional workflow nodes and approximately 691 MB latent upscaler weights. Capturing Refine state uses extra RAM and disk. Personal images, generated media and test logs are not included.

H3 Studio 9.2.0 — Spectrum acceleration and experimental Hybrid engine

Choose a tag to compare

@erdinoral erdinoral released this 11 Oct 20:30

H3 Studio 9.2.0

Spectrum acceleration

  • Optional Spectrum installation is included in Install and Update, pinned to commit 806fe498d574a726fecb8818917f00a4b993fd77.
  • Spectrum ON/OFF control appears below Video engine in Scene and Settings, with its own styling and Experimental badge.
  • English and Turkish labels update when the interface language changes.
  • Disabled by default. Applies to new native single-clip H3 graphs after model adapters; custom multishot sampling is unchanged.
  • Video blend weight 0.50, audio blend weight 0.0, offline smoothing replay enabled. Generated motion and speech can change.
  • A local RTX 5060 Ti 16 GB test at 1056×608, 14 steps and approximately 10 seconds took 1575.058 s without Spectrum and 966.344 s with Spectrum: 38.65% shorter total time. This is one matched pair, not a general performance guarantee. Generated media and personal reference images are not distributed.

Experimental Hybrid engine

  • H3 Hybrid selects the b25–49 checkpoint for both H3 model slots while retaining the existing encoder and VAEs.
  • Reference mode with the Hybrid checkpoint offers experimental 30-second requests. This does not establish that high-resolution 30-second generation fits a 16 GB GPU.
  • Hybrid r48 adapter uses its compatible loader. Other model engines remain selectable.
  • See docs/hybrid-h3.md for model sources, verification and limitations.

Validation and updating

180 Python tests and 32 JavaScript tests passed. Launcher syntax and whitespace checks passed.

Run Update in Pinokio, stop and start Studio, then reload the page. Existing local settings and generated media are retained. Spectrum installation requires a ComfyUI restart before enabling it.

Sliding-window long-video generation is not included in this release.

H3 Studio 9.1.1 — Director fixes, short clips and support

Choose a tag to compare

@erdinoral erdinoral released this 11 Oct 14:10

H3 Studio 9.1.1

Director production

  • Empty chapters now include an Add scene button that places the new scene in that chapter.
  • Produce all queues eligible scenes across chapters; Produce chapter queues only its chapter.
  • Empty production controls stay disabled, with clearer scene/chapter/all labels.
  • Chapter action buttons share the same height and alignment.

Short clips

  • Scene and individual Director scenes accept 2- and 3-second clips.
  • H3 renders its four-second frame schedule, then FFmpeg trims the video and optional audio to the requested duration. Generation time is therefore similar to a four-second clip.
  • Continuations use the last frame of the trimmed clip.

Support

  • The Support panel links to Erdin's shared Buy Me a Coffee page.
  • Open Collective and GitHub Sponsors payment buttons have been removed.
  • Star the repository is a styled button. Bug and idea reporting remain available.

Updating

Run Update in Pinokio, stop and start Studio, then reload the browser.
Existing projects and generated media remain in place.

Validation: six Director JavaScript checks and ten Python checks covering short clips, production validation and character identity. Short-clip checks include real FFmpeg trimming with and without audio.

Experimental Hybrid model and 30-second generation changes are not included in this release.

H3 Studio 9.1 — Appearance, Scene audio and creator tools

Choose a tag to compare

@erdinoral erdinoral released this 10 Oct 15:29

H3 Studio 9.1.0

The new Appearance controls are now included in the public update. Open Settings → Appearance to select a color palette and interface style independently.

Appearance and guidance

  • Orange, Blue and Graphite palettes.
  • Modern, Old School and Studio styles change typography, corners and controls.
  • Preferences persist in the current browser.
  • A bilingual in-app Guide explains Scene audio, Director, references and final assembly.

Audio, references and assets

  • Scene supports a single uploaded audio clip as a generation reference or as the original soundtrack attached to the output. Original-audio attachment does not create lip synchronization.
  • Character and asset selection fixes keep only the intended cast and references in each shot; matching library entries can be reused rather than generated again.
  • The H3/Qwen asset switch separates video-derived stills from Qwen image generation. Referenced Qwen views are generated individually through Image Edit.
  • Selecting a source image stages it for preview; generation waits for an explicit action. Deletion and polling fixes prevent removed images from reappearing.
  • Chapter production validates its queue before generating assets, with clearer continuation and scene-edit handling.

LoRAs and optional tools

  • Categorized adapters, local Civitai key entry and validated optional downloads.
  • Selected character adapters and trigger words follow the intended scene cast in Scene and Director.
  • New LightX2V FL2V/Ref2V speed adapters apply compatible step, strength and sampler recipes automatically. Only one speed recipe can be selected at a time.
  • Experimental HyperFlow uses its required model patch and sampler schedule rather than an ordinary LoRA loader.
  • A shared RefMod switch in Scene, Director and Settings enables cached full visual references. It does not train a LoRA or guarantee identity preservation.
  • Optional Look Sheets and native AV continuation workflows are available through Film Tools. Studio's existing Continue action is not automatically replaced by those separate workflows.
  • Model storage provides guarded deletion of individual managed weight files. Active/default files and queued-job dependencies are protected.

Update

Use Update in Pinokio, then stop/start Studio and reload its page. Tool installation does not automatically download the optional adapter weights. Download desired adapters through Settings or the optional LoRA installer.

Speech identity, pronunciation, lip sync and image identity remain model-dependent. HyperFlow and the optional advanced workflows need validation on each setup. Automatic transcription and burned-in subtitles are not included in this release.

Validation: 164 Python tests and 27 JavaScript tests passed. Launcher/frontend syntax checks passed. RefMod with LightX2V and experimental HyperFlow completed local video smoke tests; this does not establish comparative output quality.

H3 Studio 9.0.0 — Film & AI Director

Choose a tag to compare

@erdinoral erdinoral released this 04 Oct 07:35

MiniMax H3 Studio v9.0.0 — Film & AI Director

Create an editable film plan inside Studio, keep your cast and selected adapters attached to it, and use explicit references in both Scene and Director.

Highlights

  • AI Director turns a story and target duration into a validated film package using your configured LLM. It adds new scenes in a separate chapter, preserves existing cards and LoRA/voice assignments, and leaves production under your control. JSON import/export remains supported.
  • Shared character, location, creature and vehicle reference selection in Scene and Director. Continuations keep the previous final frame first in the reference graph; standalone Scene only auto-matches film assets when enabled.
  • Character LoRA mode skips character sheet generation and excludes the card's old images while enabled. Name and voice remain editable.
  • Categorized LoRA selectors, available trigger metadata, usage guidance and independent weights for up to three selected adapters. Existing catalog downloads are reused.
  • Optional H3 Singularity main-model switching in Settings. Approximately 21 GB, separate download, experimental; not installed by default. The switch affects future jobs and retains the encoder/VAEs.
  • Four vehicle views: front, rear, right and left. Improved individual asset-image deletion and gallery navigation.
  • Shot review with read-only repair previews before applying changes; reference video/audio excerpts without overwriting originals.
  • Wider Settings forms, responsive LoRA cards and expanded TR/EN localization.
  • Gemini streamed text now preserves word boundaries, including exact character and location names.
  • Includes the latest main-branch H3 graph API compatibility fix.

Updating

Stop Studio, select Update in Pinokio, start again, then use Ctrl+F5 in the browser. Existing models and saved projects are retained; do not use Reset for this update. Configure a reachable LLM provider to use AI Director. Optional models and LoRAs are downloaded separately.

This release does not guarantee identical faces across every shot or uninterrupted long videos. Singularity's quality advantage has not been established. Shot review checks the written plan, not rendered video quality.

Validation

Python and JavaScript regression suites cover film imports, planning validation, continuation wiring, explicit asset manifests, LoRA selection/weights, optional models, review previews, gallery navigation and translations. Live AI planning and import were also checked using a temporary film store.