<h1 id="bruxos-do-vfx-h3-camera">Bruxos do VFX H3 Camera</h1><code>#bruxosdovfx</code><div data-youtube-video><iframe width="640" height="480" allowfullscreen="true" autoplay="false" disablekbcontrols="false" enableiframeapi="false" endtime="0" ivloadpolicy="0" loop="false" modestbranding="false" origin playlist src="https://www.youtube.com/embed/XHnYSIdoONs" start="0"></iframe></div>Visual camera planner for MiniMax H3 inside ComfyUI. You drag the camera around a 3D sphere, place keyframes on a timeline, and the node compiles that trajectory into prompts that H3 understands.It compiles prompts, not camera embeddings. There is no geometric adapter here: H3 is still free to miss the angle, timing, and scale. What this node does is write the instruction in the most precise and least ambiguous way possible, and several of its design decisions exist because the previous approach failed in specific ways.It does not call any API, download anything, or require any Python dependency beyond the standard library.<h1 id="installation">Installation</h1><pre><code>cd ComfyUI/custom_nodes
git clone https://github.com/&lt;your-username&gt;/ComfyUI-H3-Camera-Editor</code></pre>Restart ComfyUI. The node appears under <code>Bruxos do VFX/Camera H3</code> with the name Camera H3 da Bruxos do VFX.<h1 id="connections">Connections</h1>Output from this nodeConnect it to<code>compiled_prompt</code><code>compiled_prompt</code> on Text Encode H3 Edit / Generate<code>options</code><code>options</code> on Text Encode H3 Edit / Generate<code>length</code>the generation frame count<code>fps</code>the <code>fps</code> input of the video creation node<code>compiled_prompt</code> and <code>options</code> are required together. The <code>minimax_prompt</code> output is an alternative to <code>compiled_prompt</code>, never an addition — connect one or the other to the same input.Also connect your image to <code>reference_image</code>. It is the same image already feeding the H3 Edit <code>source_image</code>; when connected here, it appears in the panel and the frame's actual aspect ratio is included in the prompt.<a target="_blank" rel="ugc" href="https://github.com/user-attachments/assets/33149617-bde1-4199-ae65-078f2f3dec23">https://github.com/user-attachments/assets/33149617-bde1-4199-ae65-078f2f3dec23</a>To save the video, decode the sampler result using the H3 video VAE — not the scene coverage calibrated decoder, which expects fixed windows that an arbitrary trajectory does not have.<h1 id="the-panel">The panel</h1>Drag the purple camera around the sphere to orbit. The drag locks to the axis of the initial movement: horizontal movement orbits, vertical movement changes elevation. Release and drag again to switch axes. This exists because, without the lock, trying to make a simple orbit would unintentionally introduce elevation.<ul><li>Scroll the mouse wheel to change distance.</li><li>Drag the background to rotate the viewport without changing the trajectory.</li><li>Keyframes defines how many points the timeline has, from 2 to 24. The first one is always the original image and cannot be moved.</li><li>⟳ Pure Orbit resets the elevation of every keyframe to zero while preserving azimuth. It is the shortcut for an eye-level orbit.</li><li>Reference image loads a local file into the preview. This is only necessary when the node runs outside ComfyUI; with <code>reference_image</code> connected, the image is loaded automatically.</li></ul>The panel warns you starting at 20° of elevation, when the horizon already leaves the frame, and again from 45° onward, when the video tends to become a high-angle shot.<h1 id="&quot;tests&quot;-bar">"Tests" bar</h1>At the top of the panel, two buttons enable and disable features currently under evaluation, plus one indicator:ButtonWhat it doesExtended contractsToggles the <code>prompt_detail</code> widgetSingle angle (image)Toggles the <code>runtime_task</code> widgetloop closureRead-only indicator. Turns green when the trajectory closes a full orbitThe buttons write to the actual widgets, so the selected state is saved in the workflow and the two never disagree.<a target="_blank" rel="ugc" href="https://github.com/user-attachments/assets/9bc415d7-1746-43db-a17c-72ea9722deda">https://github.com/user-attachments/assets/9bc415d7-1746-43db-a17c-72ea9722deda</a><h1 id="widgets">Widgets</h1><h1 id="camera_trajectory">camera_trajectory</h1>The trajectory in JSON format, written by the panel. Each keyframe contains <code>time</code> (0 to 1), <code>azimuth</code> in degrees, <code>elevation</code> in degrees, and <code>distance</code> as a multiple of the initial radius. It can also be edited manually. The first keyframe must be <code>time=0, azimuth=0, elevation=0, distance=1</code>, which represents the original image.<h1 id="profile">profile</h1>124, 243, or 362 frames at 24 fps. All shot timing comes from this setting: keyframe timestamps, segment ranges, and the duration declared in the prompt. That is why <code>length</code> and <code>fps</code> are outputs — connect them instead of manually entering the same numbers in two different places.<h1 id="interpolation">interpolation</h1><code>smooth</code> or <code>linear</code>. In <code>smooth</code> mode, the camera eases into and out of the shot while maintaining a constant rate through the middle; it only stops where the rotation direction actually reverses.<h1 id="instruction">instruction</h1>Free-form text inserted once, at the end of the prompt. Write only what the node cannot know: the environment, which subject is the target when there is more than one person, or a style reference. Everything else is already generated and does not need to be repeated: scene freeze, first image as reference, locked aim, zero roll, angles, timing, and a single continuous shot without cuts.<blockquote></blockquote><h1 id="subject_framing">subject_framing</h1>How much of the frame the subject occupies in the original image. Calibrated against the actual bounding boxes from the tutorial distributed by MiniMax: a distant full-body figure measures <code>W=0.071, H=0.249</code>, while a large close-up measures <code>W=0.52, H=0.701</code>.optionwidthheightwhen to use<code>close-up</code>53%72%head and shoulders<code>medium shot</code>28%56%waist up<code>wide shot</code>9.7%34%full body at a distance<h1 id="subject_box">subject_box</h1>The subject position in the format <code>[L=0.516, T=0.148, W=0.071, H=0.249]</code>. Leaving it empty uses the entire image bounds — deliberately, without guessing a bounding box. Fill it in when the subject is significantly off-center.<h1 id="minimax_format">minimax_format</h1>The same shot expressed in four different formats for the <code>minimax_prompt</code> output:<ul><li><code>coordinate only</code> — text-based coordinate block</li><li><code>coordinate + H3 sections</code> — the same coordinates wrapped in <code>subject_definitions</code> / <code>summary</code> / <code>retention_analysis</code> / …</li><li><code>compact JSON</code> — JSON object with almost no prose</li><li><code>compact JSON (no boxes)</code> — camera parameters only, without screen-space bounding boxes</li></ul><h1 id="elevation_range">elevation_range</h1>Range of the elevation control: <code>+/-15</code>, <code>+/-30</code> (default), <code>+/-60</code>, <code>+/-89</code>. It also scales the sensitivity of vertical dragging.With the assumed field of view, the horizon already leaves the frame at around 20° — at 13°, the ground occupies 82% of the image. The old ±89 range was mostly unusable and made vertical dragging excessively sensitive. Reducing the range never rewrites a keyframe: a point at 70° remains at 70°, and the slider expands to accommodate it.<h1 id="orbit_direction">orbit_direction</h1><code>invert H3 orbit</code> or <code>same as HUD</code>. This calibrates the direction between what the panel displays and what H3 produces. It does not alter the saved trajectory.<h1 id="runtime_task">runtime_task</h1><ul><li><code>scene coverage | camera path</code> (default) — video, with duration coming from <code>profile</code>.</li><li><code>directed | new camera angle</code> — a single image from a new angle. It fixes the generation to 39 frames, ignores <code>profile</code>, completes the movement within 65% of the clip, and requests that the framing remain still for the rest, because the decoder extracts the final image from that stationary tail.</li></ul>Character sheet profiles are not offered because the upstream node raises an error when they are combined with the frame anchor used by this node.<h1 id="prompt_detail">prompt_detail</h1><ul><li><code>v15 baseline</code> (default) — outputs the prompt exactly as in the previous version.</li><li><code>extended contracts</code> — adds axis separation, frame-edge direction tests, rotation completeness, degrees per second, and parallax magnitude.</li></ul>The extended mode contains almost twice as many words. A longer prompt is not automatically better, so it is opt-in: toggle only this widget while keeping the same trajectory to compare the results.<a target="_blank" rel="ugc" href="https://github.com/user-attachments/assets/0882bfde-9f62-4a1f-9bda-7da121dbe7e2">https://github.com/user-attachments/assets/0882bfde-9f62-4a1f-9bda-7da121dbe7e2</a><h1 id="outputs">Outputs</h1><h1 id="compiled_prompt-string">compiled_prompt — STRING</h1>A prose prompt using H3 sections: <code>subject_definitions</code>, <code>summary</code>, <code>retention_analysis</code>, <code>detailed_description</code>, <code>overall_soundscape</code>, <code>non_diegetic_music</code>.<h1 id="options-h3edit_options">options — H3EDIT_OPTIONS</h1>The 13 keys read by the H3 Edit encoder. All of them are explicitly populated: if any key is missing, the upstream node falls back to its hidden legacy widgets, which may retain stale values from previously saved workflows.<code>coverage_arc_degrees</code> and <code>coverage_direction</code> are derived from the actual rotation. <code>coverage_loop_closure</code> turns on automatically when the trajectory closes — see below.<h1 id="storyboard_json-string">storyboard_json — STRING</h1>The storyboard table: frame aspect ratio, duration, raw trajectory, and each segment with its camera mode, speed curve, and start/end poses.<h1 id="info-string">info — STRING</h1>Human-readable diagnostics. Connect it to a <code>PreviewText</code>. It displays the version, active task, frame count, warnings for keyframes outside the configured range, and whether loop closure is enabled.<h1 id="minimax_prompt-string">minimax_prompt — STRING</h1>The same trajectory expressed using the format selected in <code>minimax_format</code>. An alternative to <code>compiled_prompt</code>.<h1 id="length-int-and-fps-float">length — INT and fps — FLOAT</h1>Frame count and frame rate against which the shot was timed. Connect them to the generation and video nodes. If generation runs with a different frame count, the choreography describes a scene that does not actually exist.<code>fps</code> is FLOAT because that is what ComfyUI's <code>CreateVideo</code> accepts. <code>length</code> is the frame count; keyframe timestamps use the instant of the last visible frame, <code>(length - 1) / fps</code>, so the resulting file lasts one additional frame interval.<h1 id="h3world_actions-string">h3world_actions — STRING</h1>Action schedule for <a target="_blank" rel="ugc" href="https://arxiv.org/abs/2609.01560">H3-World</a>, which encodes one text clause per video latent — 37 in a 124-frame clip.<pre><code>latent 1 [0.000s-0.139s] J the camera pans left slowly
latent 37 [4.986s-5.125s] F+L+K the camera pans right and tilts up fast</code></pre><code>W</code>, <code>A</code>, <code>S</code>, and <code>D</code> are never emitted because they move the character. The output explicitly declares its own limitations, and they are not minor details:<ul><li>Pan is not orbit. It is the camera rotating in place. Perspective does not change, nothing hidden is revealed, and the subject slides out of frame.</li><li>Distance has no key, so camera radius is discarded.</li><li>Only 124 frames is a trained horizon.</li><li><code>I</code> versus <code>K</code> is not published. The text clause is what H3-World actually encodes; the key column is only a convenience.</li></ul>This does not replace the actual integration: H3-World requires the LoRA, interval-based encoding, and directed-attention routing provided by the corresponding node package.<h1 id="loop-closure">Loop closure</h1>When the trajectory closes a full orbit — an arc of exactly 360°, with the same elevation and distance as the starting point — the node enables <code>coverage_loop_closure</code>. In the upstream implementation, this flag encodes the source image a second time and anchors the final frame to it.This is a latent anchor, not a text instruction. For a complete orbit, it is the difference between asking for the rotation and forcing it: the model cannot simply stop halfway through.trajectoryloop closure360°enabledtwo rotations (−720°)enabled355°disabled360° with changing distancedisabled360° with changing heightdisabledThe final three cases matter: if the camera ends at a different radius or height, the final frame is not the same as the first one, and forcing the source image there would conflict with the trajectory.If your rotation does not complete, close the orbit. This is the only feature here that acts outside the prompt itself.<h1 id="limitations">Limitations</h1><ul><li>This is prompt-based guidance. H3 may still miss the angle, timing, and scale, and no prompt wording can completely solve that.</li><li>Without <code>subject_box</code> filled in, the node does not know where the subject is located in the frame.</li><li>Without <code>reference_image</code> connected, coordinates are normalized to 16:9.</li><li><code>directed | new camera angle</code> outputs an image, not a video.</li><li>The H3-World schedule describes pan and tilt, which represent a different camera move from the orbit drawn in the panel.</li></ul><h1 id="credits">Credits</h1>Node by Bruxos do VFX.<code>ethanfel/ComfyUI-MiniMax-H3-Edit</code>. The motion vocabulary follows the <code>buildViewPrompt</code> implementation from MiniMax's Multi-Shot skill and the coordinate format used by the Coordinate Camera Control Designer skill. The action output implements the scheme described in <a target="_blank" rel="ugc" href="https://arxiv.org/abs/2609.01560">H3-World, arXiv:2609.01560</a>.<img src="https://i.redd.it/okud6b2umpoh1.png" alt="rich text editor image" width="1049" height="681" />

Minimax h3 Camera Control

<a target="_blank" rel="ugc" href="https://github.com/NyckM/3d-Camera-control-H3-Minimax">https://github.com/NyckM/3d-Camera-control-H3-Minimax</a>

Files with SHA256: 41622725e802b739133ef8b977ac1b114eda97f529f74956714f6bc0c5c320c8