H3 Prompt Studio 1.3.1 — Text-only Game startup

H3 Prompt Studio 1.4.1

The player should be established before the opening scene, then remembered. This update fixes the missing-identity movement dead end and the confusing recovery flow.

• New text games establish missing player appearance in the existing opening planning call, before rendering. Acceptance saves it for later moves; authored appearances and explicitly assigned portraits are preserved.
• Older unbound scenes open a picture-based character choice directly from Move. Choose one person or enter a short description; the queued move continues with the updated identity.
• Recovery preserves the original move and later queue entries. Changing the selected person leaves only the new person marked as you.
• A saved ending photo assigned to a character now survives temporary movement-frame replacement, preventing broken reference bindings.

Focused checks covered opening persistence, reference retention, picker recovery, stale selections, cancellation and character switching. The built game browser recovery scenario and production build passed. The reported saved-game Right move was rendered at 0.2 MP and eight steps, then visually reviewed: the original player moved right and nearby figures stayed in place, without an LLM call. The full suites were not rerun.

Download the ZIP and verify it with the accompanying SHA-256 file. Existing model/runtime requirements still apply. See the movement guide for the exact scope and limitations of the live check. Private game media is not included; previously published demos retain their version labels.

H3 Prompt Studio 1.3.1 — Text-only Game startup

Basic eligible Game arrows now prepare movement locally, start H3 from the saved current image and skip assistant loading and automatic ending inference. Compatible returns use an earlier saved view as the last frame without restoring old inventory or enemy state.

The new In this frame panel shows inspected people and objects, their observed positions, target actions and This is me player identification. Old positions are clearly marked stale; an explicit scan refreshes them. Review controls explain when a pending take is blocking movement.

Three actual 0.2-resolution movement renders checked Forward, Back and a correction for an unwanted moving bystander. The corrected Forward completed in about 18 seconds end to end on the local setup; both warm movement renders avoided LLM inference. Focused regressions and the scene-recovery browser scenario passed, and the production frontend was built. The full suites were not rerun for this release.

H3 still generates video rather than simulating geometry. Saved-frame conditioning and more explicit actor instructions reduce drift but cannot guarantee every render. See movement behavior and measured validation for the exact limits and checks.

The ZIP includes the built interface and source. Existing ComfyUI models and native nodes are still required. Private copied game media and machine settings are excluded; earlier public demos retain their original version labels.

H3 Prompt Studio 1.3.1 — Text-only Game startup

Fixes a text-only Game opening failing with a misleading missing-model error when ComfyUI was offline.

• Extra reference images now require explicit opt-in. Ordinary Game actions use text and existing references directly.
• Offline or unreadable ComfyUI inventories are distinguished from verified missing model components. Available alternatives never silently change the selected generator.
• Game settings include generator refresh, specific availability diagnostics, and visible saved selections.
• Legacy automatic optional-image failures can be resumed only when the saved job proves it failed before submission. Existing assistant and job evidence is retained; submitted or uncertain work is never silently discarded.

The actual reported failed turn was recovered in the installed app with no extra image job. Its 608 × 320, 3.042-second clip played in the browser and the game returned to the next-move state. This verifies the reported startup path, not every possible game action or visual outcome. The earlier scene-continuity validation report now explicitly records the missed case.

The release ZIP contains application source and the built frontend; private stories, credentials and model files are excluded. The existing v1.3.0 demonstration videos remain available in that release.

For an existing installation, preserve its data folder when updating application files. Start the configured LM Studio and ComfyUI servers, then reload the app. An image generator is optional for normal Game play; H3 video still requires ComfyUI running.

H3 Prompt Studio 1.3 — Scene continuity

H3 Prompt Studio 1.3 gives Studio and Game explicit control over who acts, who stays in place, which props exist, their colors and counts, and where they start and end. Stable object IDs and accepted game state now reach the director, with bounded recovery when a response is incomplete or inconsistent.

• Edit per-shot actor holds, prop identities and scene layout. The compiler preserves authored controls and exact dialogue across all five H3 modes.
• Keep inventory, NPC identities, doors and combat consequences across accepted turns. Known mechanical actions use one directing call; inventory needs no model.
• Use output budgets that scale with scene complexity. Fresh ending inspections assess every listed character and prop, permitting uncertainty and preserving the previous branch when inspection fails.
• Recover lost responses, cancelled takes and pending decisions without silently advancing the story.

Validation: 1,569 backend + 369 frontend + 40 ComfyUI bridge tests, two built-UI browser scenarios, real local planning checks, and actual H3 rendering. The final broader planning run passed 13/13 model cases plus a deterministic locked-door guard; camera registry follow-ups passed 2/2. Earlier failures and manual limitations remain in the published evidence.

The Linux CI workflow passed, including the backend, frontend build/tests, browser scenarios, ComfyUI bridge and distributable packaging.

Included media:

• The Coin Safety Inspector — new 30-second authored continuity demo, 0.2 MP / eight steps.
• A Very Polite Pursuit — earlier 30-second chase.
• The Extremely Important Parcel — earlier one-minute parody.
• Actual generated frames, the Studio controls screenshot and provenance records.

Research and implementation · Validation · Watch the demos

Download the ZIP and its SHA-256 sidecar, extract it, and follow the README's Windows setup instructions. The ZIP includes the built frontend and source; LM Studio, ComfyUI and model weights are installed separately. No private stories, credentials or model weights are distributed.

H3 can still make visual mistakes. The demo's exact hand staging remains imperfect, and complete inspection coverage does not make the model's visual judgements infallible. The original footage is preserved without visual corrections.

H3 Prompt Studio v1.1.0 — Studio + Game

Studio and Game, with saved story branches

Studio now keeps the editor beside the video, puts continuation directly below the saved ending, and separates another take from a new scene. Previewing history no longer changes the active story. Explicit branches and selected alternatives control playback and whole-story export.

Game is a separate workspace: choose your character, type an action or exact speech, and let the selected local assistant plan and render one response. It offers three next moves after inspecting the ending, with review, cancellation, retry and alternate takes available. Studio drafts remain available when switching modes.

This release also adds durable turn and asset jobs, recovery after lost responses, installed Z-Image-Turbo BF16/FP8 selection, stable asset roles, automatic GPU handoffs, and trimmed new-footage playback. BF16 remains the image baseline; FP8 is optional.

Live RTX 4090 checks used 0.3 MP, eight steps and about five seconds of new action. H3 execution took approximately 77–99 seconds per accepted clip, excluding assistant and asset work. Player questions and replies were independently transcribed; a generated-location cut was corrected after its first take duplicated a character. See the validation record for measurements and limits. H3 can still make identity, staging, object-handling and speech errors.

Download the application ZIP and run Setup.ps1, then Launch.ps1. Before upgrading, stop the existing Studio server and back up your data directory and separate ComfyUI outputs/continuation states. Preserve the data directory when replacing application files. Models and personal projects are not bundled. The v1.0 examples and release remain unchanged.

H3 Prompt Studio v1.0.0

H3 Prompt Studio v1.0.0

Turn your reference images and a plain-language idea into a directed H3 video, then compare takes and continue the story in the same browser workspace. Connect your own local LM Studio and ComfyUI installation.

Start here

• H3-Prompt-Studio-v1.zip — the application, frontend build, setup scripts, and reviewed public examples. Extract it into a writable folder, run Setup.ps1, then Launch.ps1. Follow the repository's ComfyUI setup guide before rendering.
• H3-Prompt-Studio-showreel.mp4 — an 88-second walkthrough using actual app captures and actual generated video, with the video examples played at normal speed and their original audio.
• The-Message-Scene-1.zip, Scene-2.zip, and Scene-3.zip — three editable example projects, each with all ten reference images. Use Saved projects → Import project in Studio.

Model weights, ComfyUI, LM Studio, attention kernels, and FFmpeg are installed separately. The project exports contain reference images and authoring settings; original run history and saved video continuation states are not included.

What is included

• Named characters, assigned clothing and objects, reference tags, and easy image replacement.
• Shot timing, camera choices, cuts or continuous action, and exact dialogue by speaker.
• Local vision-assisted prompt writing with an installed LM Studio model of your choice, plus deterministic Build without AI.
• Automatic GPU hand-off between larger prompt assistants and H3.
• Inline video playback, another-seed generation, named takes, favorites, and side-by-side comparison.
• Three editable next-story suggestions grounded in the actual ending frame and saved story history.
• Saved continuation drafts, automatic continuation context, and combination of compatible clips.
• Portable project backups, reusable templates, sketch and movement guides, and an optional ComfyUI bridge.

Tested with The Message

The demonstration uses two adult characters, two outfits, three props, and three locations. Each independently directed scene conditions on eight images; the other two locations remain prompt inspiration.

Three 0.3 MP scenes rendered at 736 × 416, eight steps, and 24 fps. Each ten-second target produced 243 frames / 10.125 seconds. ComfyUI execution took 100.658 s, 90.061 s, and 88.228 s on the tested RTX 4090 system. An independent Whisper small transcription matched all 37 intended words across six dialogue lines, ignoring punctuation and case.

The café reroll retained the exact prompt and changed both the seed and decoded video pixels (87.263 s). A Qwen3.8-27B vision assistant proposed three next actions from the café ending. The selected five-second continuation rendered in 77.741 s; combining it with the selected café take produced 13.667 seconds of playback, removing 39 repeated context frames. Combining took 20.096 s and did not run video diffusion again.

These are recorded execution times, excluding prompt preparation and setup. Small hand and object details still vary: the note timing, key transfer, and exact hand assignments do not always follow the authored instructions. The audio wording check does not establish lip-sync or speaker accuracy. See the demo review for settings and limitations.

Validation

A clean Windows setup and production build passed. The standalone release passed 632 backend tests, 222 frontend tests, and 40 bridge tests. All three portable example projects passed isolated import and compiled-prompt equality checks. Videos passed complete audio/video decoding. Public files were reviewed to exclude private projects, credentials, logs, model weights, and unrelated downloads.

This is an independent community application. External tools, models, and demo media retain their own terms. The repository owner has not selected a license for the original application source; public availability does not itself grant an open-source license.