Cinema isn't a video API — it's a full production studio. Reference-driven generation, AI scoring, voice, upscaling, and assembly, all driven by your agent.
Render each clip from up to 9 reference images — character sheets, scene backdrops, motion videos. Your characters stay consistent from the first frame to the last cut.
Music beds via ACE-Step XL and MiniMax Music 3, sound design with Stable Audio 3, and expressive TTS voices. The assembler normalises everything to broadcast loudness (LUFS).
Qwen-Image-2.1 text-to-image, instruction edits, and background removal — generate character sheets, scenes, and sprites on the same server that renders your film.
cinema_assemble does LUFS normalisation, segment concat, music bed, upscale, tail fade, and publish — your finished film lands at a public URL in a single call.
// 1. Generate a character sheet qwen_image_2_1_t2i_enhanced( "single full-body front portrait, 9:16...") → ref_sheet.png // 2. Render a clip with ref2va comfy_cinema_ref2va( prompt: "[Shot 1] ...", ref_image_1: ref_sheet.png, duration: 8) → clip_01.mp4 // 3. Assemble the film cinema_assemble(clips: [ "clip_01", "clip_02"...], music: "epic orchestral") → published URL
Pre-production, production, post-production, and distribution — the whole pipeline over MCP.
Character sheets lock identity across the film: same sheet, same reference slot, same descriptors in every slide. Anti-clone guards enforce unique appearances and explicit voice ids per speaker.
Native ref2va clip generation — video and audio together, 5–10 seconds per clip, three quality tiers (0.3/0.5/1.0 megapixels). Prompt format follows the official MiniMax H3 docs.
ACE-Step 1.5 XL turbo for background scores, MiniMax Music 3 for productions with vocals, Stable Audio 3 for SFX. Per-segment music palettes mixed at 20% under dialogue.
Expressive TTS with emotion control, CozyVoice unified narration voices, and multilingual support. Every speaker gets a distinct, stable voice id reused across all slides of the film.
Qwen-Image-2.1 with the official prompt rewriter for composition-sensitive shots, instruction-driven edits, and background removal — characters, scenes, and sprite assets from one server.
SeedVR2 3B temporal-consistent upscaling (no jitter, Hann crossfade blending) to 1080p delivery, plus fast ESRGAN per-frame upscaling for clean 2D content.
Submit clips and keep working — check queue status, poll jobs, cancel stale renders, and inspect full provenance metadata for every generated asset.
Game-ready extras: transparent sprite generation with BiRefNet matting, 6-frame flipbook motion sprites, and grid sheet assembly — all from the same toolkit.
The production pipeline from idea to published film in 4 steps.
Generate character sheets (one figure, front view, 9:16) and scene backdrops with the image studio. Read cinema_flow for the full guide and grab the story template.
Write each clip prompt yourself in the MiniMax H3 format — explicit voice ids, anti-clone guards, chained transitions. No server-side script authoring; you stay the director.
comfy_cinema_ref2va renders each 5–10s clip from your references — video and audio together. Queue it and poll; the server keeps full provenance for every render.
cinema_assemble normalises loudness, mixes the music bed, concatenates segments, upscales to 1080p, fades out, and publishes the finished film to a permanent URL.
Every Cinema capability is exposed as an MCP tool. Compatible with every MCP client.
| Tool | Description |
|---|---|
| cinema_flow | The complete production guide — MiniMax H3 prompt format, workflow rules, output organisation, and the 9-reference ref2va workflow. |
| cinema_script_template | Story JSON template structuring cast, scenes, and slides for a film. |
| comfy_cinema_ref2va | Render one MiniMax H3 clip (video + audio) from up to 9 reference images — character sheets, scenes, motion videos. |
| cinema_assemble | Assemble-only finishing: LUFS normalise → concat → music bed at 20% → upscale → tail fade → publish. No renders queued. |
| qwen_image_2_1_t2i / _enhanced / _edit / _bg_removal | Full image studio — text-to-image with the official prompt rewriter, instruction edits, and background removal. |
| ace_step_15_xl / minimax_music_3 / stable_audio_3 | Music generation — XL background scores, productions with vocals, and sound design palettes. |
| expressive_tts / cozyvoice / hindi_chatterbox | Voice acting — emotion-controlled TTS, unified narrator voices, and multilingual support. |
| seedvr2_hd_upscale / video_upscale_esrgan | Temporal-consistent diffusion upscaling to 1080p, or fast feed-forward ESRGAN for clean 2D content. |
| faceswap_insightface_exact | Exact one-click face swap across a whole video — same motion, audio, and subtitles, nothing else changes. |
| sprite_asset_generator / motion_sprite / sprite_sheet_maker | Game sprite production — transparent assets, 6-frame motion flipbooks, and grid sheets. |
| kling3_omni_i2v | Image-to-video generation for B-roll and establishing shots. |
| get_job / cancel_job / get_queue_status | Async job control — queue visibility, websocket polling, and cancellation for long renders. |
| publish_asset / view_image / get_asset_metadata | Distribution — publish finished assets to public URLs, preview inline, and inspect full provenance. |
From a blank page to a published film — every stage has dedicated tools.
Story template, character sheets, scene backdrops, sprite assets.
MiniMax H3 ref2va clips, image-to-video B-roll, queued rendering.
SeedVR2 upscaling, face swap, LUFS loudness, music mixing, fades.
One-call publish to a permanent, no-auth public URL.
Cinema runs on our ComfyUI render farm at mcp-services.arcflameinteractive.com. Get access, configure your client, and roll camera.
// ~/.config/opencode/opencode.jsonc { "mcp": { "arcflame-cinema": { "type": "remote", "url": "https://mcp-services.arcflameinteractive.com/comfy/mcp", "headers": { "Authorization": "Bearer <your arcflame token>" }, "enabled": true } } }
// ~/.claude/settings.json { "mcpServers": { "arcflame-cinema": { "type": "remote", "url": "https://mcp-services.arcflameinteractive.com/comfy/mcp", "headers": { "Authorization": "Bearer <your arcflame token>" }, "enabled": true } } }
// .cursor/mcp.json { "mcpServers": { "arcflame-cinema": { "type": "remote", "url": "https://mcp-services.arcflameinteractive.com/comfy/mcp", "headers": { "Authorization": "Bearer <your arcflame token>" }, "enabled": true } } }
# Bridge any local client to the remote MCP npx mcp-remote https://mcp-services.arcflameinteractive.com/comfy/mcp \ --header "Authorization: Bearer <your arcflame token>"
Compatible with Claude Desktop, opencode, Cursor, Windsurf, and any MCP client. Access is issued with your Arcflame token.
Request AccessCharacter sheets, scripted clips, AI scores, and one-call publishing — the full studio is one MCP connection away.