ComfyUI-Cine-con-IA
Cinematic ComfyUI video nodes · Nodos cinematográficos para video con IA
Nodes (14)
You pick "contrapicado, 85 mm" and it writes the camera line H3 actually obeys
The Loader That Does the Chunking So Your Card Doesn't Die
8 Seconds Isn't 192 Frames, and Your Video Model Knows It
The Second Pass That Either Fits or Tells You Why It Won't
Reference Plates, a Locked First Frame, and That Second Conditioning Output
A sampler that refuses to run without a config wire — and that's the point
Six profiles, a traffic light, and an honest admission that none of it is benchmarked
One Button, 40 GB, and the Right Folder at the End of It
Six Sections, Fourteen Camera Recipes, and One Prompt That Actually Moves
The Node That Stops You Typing 736 Into a Width Box
Five Sampler Nodes Collapsed Into One Panel
Decode, Interpolate, Ship It — the Last Node Before Save Video
Eight boxes instead of one giant prompt — and why that isn't busywork
One text box, two outputs, and the second one is the sneaky useful one
[!TIP] 💬 Join the Cine con IA community on Discord (Spanish-speaking): questions, bugs, ideas and your own renders. discord.gg/hXKJ78cEua · Tutorials on YouTube.
Custom nodes that simplify cinematic AI video workflows in ComfyUI. The display names are deliberately generic so the pack can grow and work with several models. Its first complete flow integrates MiniMax H3 — preparation, prompt, loading, generation, refinement and output — while the Prompt node offers dedicated tabs for MiniMax H3, LTX-2.5, Wan 2.2, HunyuanVideo 1.5, CogVideoX 1.5, Mochi 1 and a model-agnostic Free mode.
The interface is in Spanish and adds visual controls, memory warnings, render progress, contextual help and camera planning tools, without turning the workflow into a tangle of technical nodes.
[!IMPORTANT] This repository contains the nodes and their interface. It does not include ComfyUI, models, LoRAs, VAEs or interpolation and upscaling weights. The Models node downloads them into the right folder with one button.
Experimental: native Acc/PDD loader
The existing Load Model node can apply the original Alibaba H3 Acc/PDD checkpoint without Deno. Full and pruned/INT8 layouts are implemented; pruned uses two bundled ~11 MB AdaLN reference grids (provenance). These auxiliary grids are the exception to the weights exclusion above; the model and Acc LoRA remain separate downloads.
Start with workflow 050: select a reference image, keep 8 steps / Euler / Simple / CFG 1 / shift 12/3, and disable refinement and progressive sampling. CPU checks with the official Ref2VA checkpoint passed; complete GPU renders, quality, speed and VRAM are not yet validated. See usage, evidence and limitations.
[!CAUTION] The four H3 nodes (Scene/Prompt H3, Camera Director H3, H3 Optimizer and H3 Optimized Sampler) are experimental. Their VRAM traffic light is calibrated with real renders on a 16 GB RTX 4060 Ti; 8, 12, 24 and 32 GB cards have no benchmark yet. See the H3 Optimizer v1 notes (Spanish).
Included nodes
| Node | Purpose |
| --- | --- |
| Cine con IA · Proporción y Tamaño | Calculates width and height from cinema, social-media, or photography aspect ratios. Supports megapixels or a fixed longest side and aligns the result to the multiple required by the model. |
| Cine con IA · Duración | Converts seconds and FPS into a valid frame count. Includes the MiniMax H3 frame grid and advanced settings for other models. |
| Cine con IA · Prompt | Builds and parses model-specific prompts for MiniMax H3, LTX-2.5, Wan 2.2, HunyuanVideo 1.5, CogVideoX 1.5, Mochi 1, or any model through Free mode. |
| Cine con IA · Cargar modelo | Loads the model, text encoder, and video/audio VAEs, as safetensors or GGUF. Chains up to four LoRAs and applies VRAM optimizations, sigma shift, and live preview when available. |
| Cine con IA · Escena | Creates H3 conditioning and the audiovisual latent. Accepts up to three reference images and a guide image anchored to a selected frame. A prompt line that starts with <Picture N> is left out when that image is not connected. |
| Cine con IA · Render | Runs the first sampling pass with direct controls for steps, sampler, scheduler, seed, and denoise. |
| Cine con IA · Escalar y Refinar | Upscales the video latent with a 3D upscaler and performs a second refinement pass. Includes 3-, 4-, and 5-step profiles and clear out-of-VRAM messages. |
| Cine con IA · Salida | Decodes video and audio, optionally interpolates frames, and returns a VIDEO object, frames, audio, FPS, and result information. |
| Cine con IA · Modelos (Models) | Shows which files each model family needs, marks the ones already on disk, and downloads the rest straight into the right folder under models/, with progress and resume. |
| CineConIA · Scene / Prompt H3 | Separates characters, action, setting, style, and sound into a structured H3 scene. |
| CineConIA · Camera Director H3 | Adds framing, angle, movement, lens, depth of field, or a complete H3 recipe and compiles the final six-section prompt. With Texto de cámara automático (experimental) each button writes the full camera text into the instruction box and the scene prompt carries no camera: see the scene prompt law and the Claude and ChatGPT skills (Spanish). |
| CineConIA · H3 Optimizer | Detects VRAM, applies AUTO/8/12/16/24/32 GB profiles, runs the Memory Planner, and returns a reusable configuration. Optional progressive sampling for larger sizes (needs the comfyui-SelfLift pack). |
| CineConIA · H3 Optimized Sampler | Consumes the Optimizer configuration and runs the first pass through ComfyUI core advanced-sampling nodes. |
| Cine con IA · Cronómetro | Browser-only node, not part of the render: times the whole run and each node, and keeps the last 20 runs in the workflow with the model (from any loader), LoRAs, steps, sampler, progressive, refine pass, seed, final resolution and duration. Click a run to see its per-node breakdown; Copiar tabla copies the history for a spreadsheet. Known models and LoRAs are logged under their official names, and every run is also added to a permanent log (ComfyUI/user/default/cineconia/registro_cronometro.csv) that Descargar registro downloads. |
| Cine con IA · Interruptor | Browser-only node, never sent to the server: each button turns some nodes on and the nodes of the other buttons off. Buttons come from the groups whose title starts with a prefix ("RAMA" by default) or are created with + botón from the selected nodes. Right-click to rename, recolor, reorder or hide them, and to pick the mode (always one, one or none, several) and whether off means bypass or mute. |
| Cine con IA · Selector | Browser-only node, never sent to the server: independent rows of buttons (for example MODEL and STEPS) that set values on other nodes or turn a node on or off, so options combine without duplicating branches. It names the video after what is selected. + at the end of a row saves the current values as a new button; right-click builds rows from the controls of the selected nodes and renames, recolors, updates or deletes each button. |
Interface highlights
- Quick controls for aspect ratio, resolution, duration, FPS, scale, and sampling parameters.
- Live information about final resolution, megapixels, relative cost, real duration, and the recommended H3 range.
- Statistical progress panels for both passes, with a real step-time chart, last/average step time, percentage, and ETA.
- Up to four LoRAs applied as an ordered chain.
- Responsive prompt tabs for MiniMax H3, LTX-2.5, Wan 2.2, Hunyuan 1.5, CogVideoX 1.5, Mochi 1, and Free.
- Shot size, angle, and movement selectors with automatic English phrasing.
- Shot history saved inside the workflow to help vary camera coverage.
- Buttons to copy a source-based recipe for an AI assistant, paste its response, and automatically distribute each model's fields.
- Compatibility with workflows saved under earlier node display names.
Model support and naming
The names shown in ComfyUI are generic: Load Model, Scene, Render, Upscale & Refine, and Output. This is intentional and leaves room for additional model backends without changing the workflow vocabulary.
The current full generation pipeline is implemented for MiniMax H3. Prompt preparation is model-independent and also includes LTX-2.5, Wan 2.2, HunyuanVideo 1.5, CogVideoX 1.5, Mochi 1, and Free modes. These additional tabs create positive and negative text that can be connected to the corresponding ComfyUI workflow; they do not replace that model's loader, conditioning, or sampler nodes.
Compatibility is checked against current ComfyUI support before a named model is added. ComfyUI lists native video support for Wan 2.2, LTX-Video, HunyuanVideo 1.5, CogVideoX, Mochi, and MiniMax H3. Keep ComfyUI updated because model support and workflow templates evolve over time.
Some internal identifiers still end in H3, such as CineCargarH3, CineEscenaH3, and CineRenderH3. These identifiers are hidden from normal use and are preserved exclusively for backward compatibility: changing them would break existing saved workflows.
Requirements
- A recent ComfyUI installation with the native MiniMax H3 nodes.
- The models and VAEs required by the MiniMax H3 workflow.
- ComfyUI-KJNodes, recommended for attention/FFN chunking and live preview. The loader continues without these optimizations when KJNodes is unavailable.
- Comfyui Minimax H3 Latent Upscaler, required only by Escalar y Refinar.
- A GGUF loader that supports H3, such as ComfyUI-GGUF-Loader, only if you use
.ggufmodels (workflow 049). - A compatible frame-interpolation model when interpolation is enabled in Salida.
The nodes only use Python dependencies already supplied by ComfyUI; this package installs no additional Python libraries.
Installation
Git (recommended)
Open a terminal in ComfyUI/custom_nodes and run:
git clone -b comfyui --single-branch https://github.com/chaLords/ComfyUI-Cine-con-IA.git
Restart ComfyUI and look for the Cine con IA category.
The comfyui branch carries only the node: the code, the interface, and what it needs to run. The example workflows, documentation, and tests stay on main, here on GitHub.
Manual installation
- Download the ZIP of the latest release (under Assets, Source code (zip)). It contains only the node.
- Extract it inside
ComfyUI/custom_nodes. - Confirm that
__init__.pysits directly inside that folder, for exampleComfyUI/custom_nodes/ComfyUI-Cine-con-IA/__init__.py. - Restart ComfyUI and look for the Cine con IA category.
Updating
If you installed with Git, open a terminal in ComfyUI/custom_nodes/ComfyUI-Cine-con-IA and run:
git pull
If you cloned it earlier without -b comfyui, you have the whole project: git pull keeps working, but to keep only the node, delete the folder and clone it again with the command above.
If you installed from the ZIP, delete the folder and extract the new release. Either way, restart ComfyUI. What changed in each version is listed in Releases and in the changelog.
ComfyUI-Manager and Comfy Registry
The package is registered in the Comfy Registry as cine-con-ia, but its versions are still pending review, so the Manager may not list it when you search for Cine con IA. Until then, install it with Git or from the ZIP. The registry package also contains only the node. Once the registry approves it, it will also install with:
comfy node install cine-con-ia
Skills for writing the prompt
The skills write the scene prompt for you. Attach your character sheet, the objects, and the location, and write only the action. You get back the prompt, ready to paste into the Prompt node, and which image goes into each reference input of the Escena node. The camera comes from the Director's buttons. The skills reply in Spanish; the prompt is in English.
<p align="center"> <a href="https://github.com/chaLords/ComfyUI-Cine-con-IA/releases/latest/download/cineconia-escena-h3.zip"><img alt="Download the skill for Claude" src="https://img.shields.io/badge/Download-Skill%20for%20Claude-D97757?style=for-the-badge"></a> <a href="https://github.com/chaLords/ComfyUI-Cine-con-IA/releases/latest/download/cineconia-escena-h3-chatgpt.zip"><img alt="Download the skill for ChatGPT" src="https://img.shields.io/badge/Download-Skill%20for%20ChatGPT-10A37F?style=for-the-badge"></a> </p>- Claude: on claude.ai, Settings → Capabilities → Skills → "Upload skill", and choose
cineconia-escena-h3.zipas it is, without extracting it. - ChatGPT: extract the zip. In Explore GPTs → Create → Configure, paste
INSTRUCCIONES_GPT.mdinto Instructions and upload the two files inconocimiento/under Knowledge. - Paste the prompt into the Prompt node of your Cine con IA workflow and connect the images as it tells you. Workflow 051 already has the Director with Texto de cámara automático.
More details in skills/README_ES.md (Spanish).
Recommended workflow
Aspect Ratio & Size ─┐
Duration ────────────┼─> Scene ─> Render ─> Upscale & Refine ─> Output ─> Save Video
Prompt ──────────────┤ ▲ ▲ ▲
Load Model ──────────┘ └─────────┴──────────────┘
Important connections:
- Connect
positivefrom Escena to Render. - Connect
positive_escalarfrom Escena to Escalar y Refinar. This output omits the first-resolution guide-image anchor and prevents shape incompatibilities during the second pass. - Connect the first-pass
latentto Escalar y Refinar, or disable that node for quick tests. - Connect
videofrom Salida to a Save Video node.
Using the Prompt node
MiniMax H3
The prompt is organized into six sections:
subject_definitionssummaryretention_analysisdetailed_descriptionoverall_soundscapenon_diegetic_music
The camera selector can replace an existing shot instruction or insert a new one into detailed_description. The H3 tab also offers the 14 named shots verified by LoopForge. A recipe writes the shot's opening framing, the subject action it needs, and the camera clause from LoopForge's published prompt word for word; pronouns follow <Subject 1> in subject_definitions. Where LoopForge's clause names its own scene, the recipe uses the generalised version from its shot recipes and leaves {SLOTS} such as {THE_SPACE} for you to fill; the prompt is not built while a slot is empty. When you paste an AI-written camera route, or select a named recipe, its complete route is preserved in the final prompt. The simple framing/angle/motion buttons remain available for a free-form shot; Change shot in text intentionally replaces the recipe with those simple controls. Check View final prompt before rendering.
The short way: pick a recipe and press “Armar el prompt con la receta elegida”. A window asks for what the recipe cannot know — who is in the shot, where they are, what is behind them and what changes in them — and writes all six sections in MiniMax's format: the subject definition with its identity lock, the summary with [reference generation], the retention analysis, the style line and the shot, the soundscape and the music. The recipe supplies the camera, the opening framing and the action it was verified with; you supply the scene. When a recipe needs a second character or a detail of its own, it asks in the same window, with an example inside each field.
The recipes only move the camera under LoopForge's test conditions, and several of them are load-bearing:
- One identity plate per character and no background plate. Describe the place in text. LoopForge measured that a background plate roughly halves the camera move.
<Picture N>is positional:referencia_1is<Picture 1>whatever the text says. [reference generation], never[keyframe completion], and never call the plate the first frame.- 124 frames (5.17 s at 24 fps) for single moves; 192 (8 s) for yo-yo zoom and split screen.
- 20 steps,
res_multistep/simple, turbo LoRA off. A 4-step turbo render suppresses camera motion almost entirely. - No timestamps or waypoints. H3 honours the order of events, not their timing.
- A person who must stay put is written standing in place, breathing softly…, never completely still: an impossible pose cost the orbit about 80% of its travel.
Whip pan needs a second character plate, rack focus a second and third; handheld, snorricam and crane rise need the subject walking or running. Every recipe was re-run here on a different character and different locations: the results, and what length and aspect ratio each one needs, are in the recipe tests. These are prompting recipes, not guaranteed physical camera controls. Judge an orbit on its stills: facing camera → back to camera → facing camera. LoopForge notes that its end background does not exactly match the start.
For image-guided H3 shots, Scene provides two guide modes:
exact · locks frame 0preserves the connected guide as an exact latent keyframe. Use it when the opening frame must match precisely.flexible · prioritizes camerauses the image as a visual reference without the exact latent anchor. Use it to test large arcs, trucks, pedestals, and other viewpoint changes. The image is then added as one more reference, after the connected ones, so it takes the next<Picture N>; if it shows the set, it acts as a background plate and halves the camera move.
Only an image explicitly used as the first frame fixes the composition at 0.00 s; a character reference plate does not. With a moving camera, identity and scene geometry remain consistent while viewpoint and parallax can change. A complete 360° orbit or yo-yo zoom may return to the opening composition, whereas a delayed crash zoom need not begin immediately. wide and slow remains available separately because a large slow arc is not equivalent to a fast marked move.
LTX-2.5
Produces a single continuous paragraph and adapts camera terminology to LTX vocabulary. The audio field is appended to the same prompt.
Workflow 049: MiniMax H3 with GGUF Q4
The 049 keeps the 048 and adds GGUF Q4 to the Selector, next to the official model and Singularity. It starts at 8 draft steps and keeps the references, the refine pass, the output with audio and the Cronómetro.
Save minimax_h3_ref2va_pruned-Q4_K.gguf in models/diffusion_models/minimax/. Cargar modelo detects .gguf and hands it to the registered GGUF nodes, also for a GGUF text encoder if you pick one. The 049 keeps the text encoder and VAEs of the 048. H3 needs a compatible loader such as ComfyUI-GGUF-Loader; the city96 version that was checked does not recognize its architecture. Install only one implementation of the UnetLoaderGGUF and CLIPLoaderGGUF nodes to avoid clashes, and restart ComfyUI after installing it.
GGUF Q4 makes the model smaller but does not guarantee a faster render: that depends on the dequantization cost and on how many RAM-to-GPU transfers it saves. Compare with the same scene, seed and steps; there are no measured times for this example yet.
Wan 2.2, HunyuanVideo 1.5, CogVideoX 1.5, and Mochi 1
Each model has its own tab and its own AI-assistant recipe. The intended workflow is:
- Copy the instruction from the selected tab and paste it into an AI assistant.
- Answer its questions about the shot.
- Paste the returned labeled block into the node.
- Review the separated fields. The node joins them in model-specific order and emits
promptandnegativeas independent outputs.
The fields are an editing surface, not a new syntax imposed on the model. Wan emphasizes motion and camera continuity; Hunyuan follows its documented component order; CogVideoX uses a detailed temporal caption within its 224-token encoder limit; Mochi favors concrete photorealistic motion.
Free
Joins two fields using a configurable separator without rewriting their contents. This mode supports current or future models that use a different prompt format.
Official model and ComfyUI references
Prompt recipes are based on the model authors' documentation, with ComfyUI compatibility checked separately:
- MiniMax H3 model card and ComfyUI package/workflows
- LTX-2.5 model card
- Wan 2.2 I2V, Wan 2.2 T2V, and official ComfyUI examples
- HunyuanVideo 1.5 model card
- CogVideoX 1.5 T2V and CogVideoX 1.5 I2V
- Mochi 1 model card and official ComfyUI example
- ComfyUI repository and native model support list
Models and files
The selectors read directly from the folders configured by ComfyUI:
models/diffusion_models: diffusion model.models/text_encoders: text encoder.models/vae: video VAE and audio VAE.models/loras: optional LoRAs.models/vae_approx: small VAE for live preview.models/latent_upscale_models: 3D latent upscaler.models/frame_interpolation: frame-interpolation model.
Exact filenames depend on the models installed on your system and appear automatically in each selector.
For MiniMax H3, the Models node also offers the optional
Minimax-h3_Singularity_ref2va_Pruned_v1.3_int8.safetensors checkpoint. It is
a complete Ref2VA checkpoint (about 21 GB), not a LoRA, and is downloaded from
the Singularity repository
into models/diffusion_models. It remains optional and is never downloaded
automatically. After downloading it, refresh the model lists or restart ComfyUI,
then select it in Load Model while keeping the MiniMax H3 profile.
Memory and performance
Video generation uses a significant amount of VRAM. The loader can split attention and FFN processing into chunks, reducing peak memory usage at the cost of speed. Refinement cost grows approximately with the square of the scale: for example, x2 processes close to four times the first-pass area.
If refinement does not fit in memory, try the following in order:
- Reduce the scale, for example from
1.7to1.5. - Reduce the first-pass megapixel setting.
- Increase attention or FFN chunking.
- Temporarily disable upscaling and keep the first pass.
Privacy
The nodes contain no telemetry or tracking. The only network access is the Models node: when you press its download button, it fetches the files you picked from their public Hugging Face repositories into models/. Nothing is downloaded on its own, and everything else runs inside your local ComfyUI installation.
Development
Run the syntax and standalone-function checks with:
python -m compileall -q .
python -m unittest discover -s tests -v
Work on main, which has the tests, documentation, and workflows: git clone https://github.com/chaLords/ComfyUI-Cine-con-IA.git. What users receive is listed in tools/paquete.py (NODO). Every push to main copies that list to the comfyui branch, and every release attaches the skill packages.
Internal node identifiers (CineCargarH3, CineEscenaH3, and the others) must remain stable to preserve compatibility with saved workflows; they are not the names displayed in ComfyUI.
Project status
This project is under active development. Keep a backup of important workflows before updating.
License
Released under the MIT License. You may use, modify, and redistribute the code as long as the copyright notice and license are retained.
The fourteen MiniMax H3 camera recipes reproduce Loop Forge's published camera clauses word for word. That work is MIT licensed, Copyright (c) 2026 Loop Forge, and its notice and licence text are kept in THIRD_PARTY_NOTICES.md along with the MiniMax guides this node's prompt format follows.