Video

Quick Create, Digital Human, Image-to-Video, Motion Transfer, and Custom Assets. Character Library is under Resources; Storyboard Templates are under Settings.

Sidebar order: Quick Create → Digital Human → Image-to-Video → Motion Transfer → Custom Assets.

Which Pack to Pick on Each Page

Sidebar feature	Dropdowns on the page	Pack capability	Built-in packs
Quick Create · Generate Images	Image provider + image workflow	t2i	qwen_image_21 / qwen_image_21_uncensored, flux3_klein, t2i_basic
Quick Create · Generate Video	Video provider + video workflow	Defaults to t2v; i2v is also available after picking a character	minimax_h3_t2v (official), minimax_h3_t2v_uncensored (uncensored, no enhancement), minimax_h3_t2v_enhancer (uncensored + enhancement); after picking a character, the i2v family appears too (first frame, multi-reference, orbit; the long-video Extender is mainly for the Director)
Digital Human	Video provider + pack; pick a voice separately	avatar, etc.	minimax_h3_avatar / minimax_h3_avatar_uncensored
Image-to-Video	Video provider + workflow pack	i2v	minimax_h3_i2v (first frame; the pack also has an optional last-frame slot) / minimax_h3_i2v_uncensored / minimax_h3_i2v_enhancer; also minimax_h3_i2v_ref (multi-reference), minimax_h3_orbit (360° orbit), and minimax_h3_extender (long video, mainly used by the Director)
Motion Transfer	Video provider + workflow pack	motion, etc.	minimax_h3_motion / minimax_h3_motion_uncensored
Custom Assets · Asset Animation	Animation provider + animation workflow	i2v	minimax_h3_i2v / minimax_h3_i2v_uncensored / minimax_h3_i2v_enhancer and other i2v packs
Custom Assets · Write Copy from Assets	Analysis provider + analysis workflow	analysis	analysis_image_qwenvl / analysis_video_qwenvl
Character Library · Generate from Description	Not selected on this page. Uses Settings → System Settings → Default Character Generation	t2i / i2i	Pick three packs: text-to-image, single-reference, and multi-reference; defaults are qwen_image_21, qwen_image_21_edit, qwen_image_21_edit_multi

The sidebar's Tools → Video Enhance is a separate page. It uses a video provider and a v2v workflow pack (built in: seedvr2_3b_video_restore) and does not depend on the video generation pages above.

How to Choose on the Page (General)

1. Video / image / animation provider: an enabled provider with a matching capability. After choosing ComfyUI, the model list is fetched from that machine's /models/....
2. Workflow pack: filtered by the capability in the table above. The dropdown first shows the capability name and a one-line description; uncensored packs are marked "Uncensored". The bulb on the right opens the full description. After selecting a pack:
   • Models declared by the pack but missing on disk → yellow warning (submission is not blocked).
   • Expand the parameters: main model, text encoder, VAE, LoRA, and so on.
   • If the pack has accel → a Turbo acceleration or Uncensored LoRA toggle. MiniMax H3 Turbo carries sampling presets: turning it on sets steps and sampler to 8 steps and euler; turning it off restores the pack's own defaults. Values you already filled in for those two fields are overwritten too. The image uncensored LoRA has no step preset; the toggle only attaches or removes the LoRA. Currently MiniMax H3 video, Digital Human, Motion Transfer, and Qwen image-edit packs enable acceleration by default; turn it off in the parameters when you want to compare quality.
3. Width/height and duration: Image-to-Video uses an aspect ratio snapped to the pack's multiple (usually 32 for H3). Packs with fps take a duration in seconds; on submit it is converted to frames and snapped with align_base / align_offset (H3 uses 17k+5). Digital Human does not show a duration; the picture follows the voice.
4. H3 / long jobs: set ComfyUI timeout in Settings → System Settings → Run Parameters to at least 3600 seconds.

MiniMax H3: Version = Complete Set of Files (Do Not Mix)

One "feature" on Comfy often maps to multiple files. The dropdowns let you change file names individually, but the official / uncensored / quantization spec must be chosen as a complete set; otherwise loading may fail or the result may be wrong.

Shared Files (Required by Official and Uncensored)

Role	Directory	Current recommended files (common remotely)
Main model FL2VA (first frame / first-last frame, text-to-video)	diffusion_models/	minimax_h3_fl2va_pruned_int8_convrot.safetensors
Main model Ref2VA (reference image / reference video, long video)	diffusion_models/	minimax_h3_ref2va_pruned_int8_convrot.safetensors
Video VAE	vae/	minimax_h3_video_vae_fp16.safetensors
Audio VAE	vae/	minimax_h3_audio_vae_fp32.safetensors
Turbo LoRA (optional)	loras/	minimax_h3_turbo_v4_step600_ema.safetensors. The node is MiniMaxH3TurboLoRA (requires Larryvrh/ComfyUI-MiniMax-H3-Turbo); keep the scheduler at simple. The dropdown only lists weights whose filename contains _ema

FL2VA and Ref2VA are two separate sets of weights. Do not mix them in one workflow: minimax_h3_i2v, minimax_h3_orbit, minimax_h3_t2v, and Digital Human use FL2VA; minimax_h3_i2v_ref, minimax_h3_motion, and minimax_h3_extender use Ref2VA. minimax_h3_orbit uses minimax_h3_video_vae_int8_convrot.safetensors as its default video VAE, while other H3 packs default to fp16; it also uses loras/minimax_h3_flf2v_lora_v1.safetensors (a first-last-frame orbit LoRA), not Turbo. When you change the main model, the whole chain still uses the same VAE. Other quantization variants (unpruned INT8, FP8, W6A8, etc.) have different sizes; do not mix them.

RTX 3090 / Ampere: use Comfy-Org's minimax_h3_ref2va_pruned_int8_convrot.safetensors (about 21GB). Community NVFP4 weights target Blackwell; on non-Blackwell cards NVFP4 is emulated, so speed and compatibility are not guaranteed.

Text Side: Official Set vs Uncensored Set

Set	Text encoder (TE, text_encoders/)	Generation tail (tail, only needed for prompt enhancement)	Matching built-in packs
Official	qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors	Usually not needed; official enhancement needs a matching …instruct_generation_tail…	minimax_h3_i2v / minimax_h3_t2v
Uncensored NVFP4 (no enhancement)	qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors	Not needed	minimax_h3_t2v_uncensored / minimax_h3_i2v_uncensored
Uncensored INT8 + enhancement	qwen3vl_32b_h3_ultra_uncensored_heretic_int8_convrot.safetensors	qwen3vl_32b_h3_generation_tail_50_63_int8_convrot.safetensors	minimax_h3_i2v_enhancer / minimax_h3_t2v_enhancer

Pairing rules:

• TE and tail must be from the same family (official pairs with the instruct tail; uncensored pairs with the matching heretic tail).
• CLIPLoader type is minimax for all of them; changing files does not change the node type.
• In minimax_h3_i2v, swapping only the TE dropdown to an uncensored file gives you uncensored output without prompt enhancement (Guide is still optional). To "rewrite the prompt from an image", you must use an enhancer pack and install the Guide node pack.

Understanding Quantization

The same role often has several quantization files (for example TE in nvfp4 / int8_convrot / bf16). Pick one spec and use it throughout. Do not use pruned-int8 for the main model and then mix in an unverified TE combination. On a 3090 the official nvfp4 runs but is not the fastest; the uncensored TE (about 25G) uses more memory, so on a remote 62G machine watch separately for OOM.

A more detailed file list and download links are in Deployment Resources.

Custom Assets

You already have images or videos; add voiceover and subtitles and join them into one short video in upload order. The picture stays as-is by default.

From top to bottom the page is: Assets → Copy and Title → Asset Animation → Voice → Aspect Ratio, Subtitles, and Background Music. When Asset Animation is on or a video is uploaded, the voice can be switched to model audio, keeping the animation or video's own audio track. After clicking Start Compositing, the order is: copy is finalized → if there is a title, make a 2-second title card first (no subtitles burned in) → with custom voice, synthesize each segment and write subtitles; with model audio, write subtitles using each clip's own duration → keep images still or animate them according to the toggle → join into one video → mix in background music if selected.

Assets

The order is the shot order. You can add multiple times, or click Choose from Library to pick existing images or videos. Each video segment is looped or trimmed to the duration of that voiceover line.

Copy and Title

The title is optional. If filled in, the video opens with a 2-second title card.

There are three copy modes; they do not affect each other:

Option	How the voiceover is produced	Looks at the picture?
Your own copy (one line per asset)	You write it. One line per asset, matching asset order; if there are fewer lines, they repeat	No
Generate copy from topic	Fill in only the topic and target duration. Copy is written from filenames and video durations. Requires the default LLM in System Settings	No
Write copy from assets	Upload assets, then click Generate Copy. A vision model looks at each image and video, then writes the voiceover. You can edit the result; the edited text is what gets used	Only when you click Generate Copy

Writing copy from assets also requires:

• Analysis provider: the ComfyUI that runs the analysis.
• Image analysis workflow: required when there are images. Built in: analysis_image_qwenvl, which needs ComfyUI-AILab and Qwen-VL (default Qwen3-VL-8B-Instruct).
• Video analysis workflow: required when there are videos. Built in: analysis_video_qwenvl, which also needs Video Helper Suite to extract frames.
• Extra requirements, target duration: optional. Use them to specify tone, selling points, and sentence length. Leave blank to write only from the picture.

Generate Copy is unavailable when no assets are uploaded. You cannot start compositing if this voiceover is empty.

Asset Animation

The toggle is off by default. When on, it takes effect only inside Start Compositing, after that segment's voice is ready: each still image goes through i2v once to generate a short motion clip, which is then looped or trimmed to the voiceover duration and given subtitles.

Uploaded videos and the title card stay as-is. The toggle is unavailable when the current assets contain no images.

If the animation prompt is empty, the default is a slight, natural camera push-in. If filled in, your motion description is used. Pick an i2v pack as the workflow, for example minimax_h3_i2v, minimax_h3_i2v_uncensored, or minimax_h3_i2v_enhancer; multi-reference and orbit packs can also be used, with images sent to their own image slots. If the pack has accel, Turbo can be enabled. Each image runs one image-to-video job; more images take longer.

Voice, Aspect Ratio, and Background Music

The voice is synthesized segment by segment following the assets, and subtitle timing follows that voice line. You can turn off burning in, and change subtitle font size and color: font size 0 means automatic, and the color defaults to white. The title card does not get subtitles. If the voice provider is ComfyUI, choose a speech workflow separately; the voiceover text is injected automatically. Local TTS or API TTS can be previewed under the voice, see Text-to-Speech. Aspect ratio is 9:16, 1:1, or 16:9. Background music is chosen from the Music Library and mixed in after the whole video is joined.

Digital Human

Character image + voiceover copy → TTS voice → avatar workflow pack. Two built-in lip-sync workflows are based on the MiniMax H3 joint audio-video model (no extra lip-sync plugin or audio encoder needed):

Pack	Default text side	How it works and what it is good for
minimax_h3_avatar	Official TE	MiniMaxH3ImageToVideo locks the first-frame character; MiniMaxH3AddGuide anchors the TTS voice latent at frame 0 to drive lip movement and expression, and the finished video carries the clean TTS audio directly; Turbo can be enabled
minimax_h3_avatar_uncensored	Uncensored NVFP4 TE	Same as above, but uses the uncensored text encoder by default; Turbo can be enabled

Pick a voice provider and voice separately. The picture duration is determined by this voice; you no longer enter seconds on the page. Frame count is rounded up to 17k+5 and then trimmed back to the voice length, leaving only about one extra frame. If the estimate exceeds about 12 seconds, or the result still exceeds the per-segment limit (362 frames, about 15 seconds), it is split by sentence; the next segment uses the previous segment's last frame as its first frame, and the segments are joined into one video. Later segments no longer return to the character sheet, so with many segments the character may still drift. After compositing, you can add background music from the Music Library. Local TTS or API TTS can be previewed under the voice, see Text-to-Speech.

The character image can be uploaded, chosen with Choose from Library, or picked from Character Library. When a character is selected, only the first segment uses the facial half-body close-up on the right side of the character sheet as its first frame; it does not use the full-body image or the entire sheet. Uploaded images take precedence over Character Library. If the character has a bound voice, it is used when no other voice is selected.

Image-to-Video

First frame + prompt → i2v. The minimax_h3_i2v pack also declares an optional last-frame slot, last_frame; the standalone page only sends the first frame. To use first and last frames, bind the last frame on a canvas function node. The finished video keeps the model-generated audio track; you can also add background music from the Music Library.

Pack	Default text side	Characteristics
minimax_h3_i2v	Official TE	First frame is required; the pack also has an optional last-frame slot last_frame, but the standalone page only sends the first frame; bind the last frame on a canvas function node; the prompt goes into generation as-is; Turbo can be enabled
minimax_h3_i2v_uncensored	Uncensored NVFP4 TE	Uncensored, no enhancement; first frame only; LightX2V 8-step by default; Turbo can be enabled
minimax_h3_i2v_enhancer	Uncensored INT8 TE + tail	First frame only; Guide enhancement runs before rendering; the enhanced text is visible in the result area; Turbo can be enabled; Guide is required
minimax_h3_i2v_ref	Official TE	Up to 9 reference images in the pack; the standalone page adds only one image_2 reference slot, and full multi-reference is bound on a canvas function node or in the Director; images only provide appearance and are not used as the first video frame; refer to them with <Picture N> in the prompt; Turbo can be enabled
minimax_h3_orbit	Official TE	360° orbit shot; with one image, the same image is used for both start and end, and a second image uploaded on the page becomes the ending frame; leaving the prompt empty uses the built-in orbit sentence; with two different images and the original prompt unchanged, an outfit-change sentence is added automatically; default 73 frames (about 3 seconds), 28 steps, no Turbo needed
minimax_h3_extender	Official TE	Long-video Extender: clips_json describes each segment's duration and prompt, reference images are passed through ref_1 to ref_9, and each segment is at most 15 seconds; mainly generated by the Director

The enhancer is slower and uses more memory; it is suitable when the prompt is very short or when you want it rewritten into an H3-style long prompt first. i2v_ref reference images are not the first frame; they are for giving character, clothing, and scene references and letting the model choose the starting camera position. orbit is good for orbiting a character or product in place; it is most stable when the camera and subject stay still, and it drifts easily when the scene contains a lot of motion.

After selecting a character from Character Library, choose the camera view for this shot: facial half-body close-up, full-body front, full-body side, or full-body back. The page automatically sends only one camera-view image for the character. The ending frame for orbit and the second reference slot for i2v_ref can be uploaded on the page; the last-frame slot for i2v and the third and later reference images for i2v_ref must be bound on a canvas function node. Use close-ups for near shots and the matching full-body image for full-body motion. Quick Create can rotate views across shots, which uses multiple character sheet images. Images can be uploaded or chosen from the library; uploaded images take precedence over Character Library.

Motion Transfer

Reference video + target image + prompt → motion workflow pack. The reference video can be uploaded or chosen from the library. Picture duration follows the reference video; you no longer enter seconds on the page. The reference video is split at the motion pack's length.max (the current two H3 motion packs use 124 frames, about 5.2 seconds, and satisfy 17k+5); the next segment uses the previous segment's final frame as <Picture 1>, while the side and back references are kept, and the segments are joined into one video. Audio can keep the model-generated track or be replaced with the reference video's entire original audio. Either way, you can add background music from the Music Library. Two built-in motion-transfer workflows are based on MiniMax H3 multimodal reference (MiniMaxH3ReferenceToVideo). When a character from Character Library is selected, full-body front, full-body side, and full-body back are sent as <Picture 1>, <Picture 2>, and <Picture 3>. The right-side facial close-up is not used for motion transfer, to avoid losing lower-body clothing. The target image can be uploaded or chosen from the library; uploaded images take precedence over Character Library.

Pack	Default text side	How it works and what it is good for
minimax_h3_motion	Official TE	VHS_LoadVideo extracts frames at 24fps as the <Video 1> motion reference, and the target image is the <Picture 1> identity reference; body motion and camera movement transfer without skeleton preprocessing; Turbo can be enabled
minimax_h3_motion_uncensored	Uncensored NVFP4 TE	Same as above, but uses the uncensored text encoder by default; Turbo can be enabled

In the prompt, <Picture 1> refers to the target image and <Video 1> to the reference video. If no tags are written or the prompt is empty, the pipeline adds the standard motion-transfer instruction automatically. You can also import your own WanAnimate2ToVideo or WanSCAILToVideo workflow pack.

Quick Create

Quick Create turns voiceover into a short video. The voiceover is the line spoken in each shot; the voice reads that line and the subtitle shows it. It is not a finished video file.

A finished video can have multiple shots. Each shot is made separately as a short segment: the line goes to the voice, and the picture follows that shot's visual prompt. All segments are joined in order, and background music is added if needed, to produce the mp4 you download. A shot count of 4 means 4 short segments joined into 1 video, not 4 separate videos.

The AI Create, Fixed Script, and Batch Create tabs at the top only decide where the voiceover lines come from. The storyboard template, voice, subtitles, character, and background music below are shared by all three modes.

What Is in a Finished Video

Each shot has two pieces of text:

• Voiceover: the line that is spoken, and also the subtitle for that shot.
• Visual prompt: the description handed to the image or video workflow. Static templates do not use it to generate images; the text is laid directly into the layout.

The title is one line shared by the whole video. It is displayed on the picture and is not read as one of the voiceover lines.

You can enable Pause after storyboard generation; composite after manual confirmation. When enabled, the script and storyboard JSON are produced first, and compositing starts after you edit each shot's visual prompt. Batch Create does not have this option.

AI Create

Use this when you have not written a script yet and only know what to talk about. Enter a topic, then set the shot count (1 to 12, default 4).

The original topic becomes the video title and is displayed on the picture. The default large language model (LLM) in System Settings writes the voiceover according to the shot count. With a shot count of 4, it writes 4 lines of about 20 to 40 Chinese characters each, in a conversational tone. Those 4 lines map to 4 shots: line 1 goes with picture segment 1, line 2 with segment 2, and the four segments are joined into one video. A title invented by the LLM is not used; the title is always the topic you wrote.

The ✨ on the topic field only expands your topic into something more specific, to make the later voiceover easier to write. The voiceover itself is written by the LLM only after you click Start Creating.

If the template is Generate Images or Generate Video, there is also Use the LLM to write visual prompts during generation, on by default. When on, after the 4 voiceover lines are written, the LLM rewrites each line into a visual description for image or video generation. When off, the visual description uses the voiceover text as-is. Workflow packs with Guide (names containing enhancer) also rewrite the prompt once inside the pack; packs without it (for example uncensored) do not. That in-pack enhancement and the page toggle are two separate things; turning one off does not turn off the other.

Static style templates do not write visual prompts. The voiceover text is laid directly into the layout and read aloud.

An LLM must already be configured. If it is not, the page prompts you to go to System Settings or Providers.

Fixed Script

Use this when the script is already written and you want the exact sentences spoken. The voiceover is exactly the text you paste in; the system does not rewrite it into other sentences. The topic field next to it is only the title, displayed as-is and not passed through AI.

This page has no shot count. How many shots there are is determined by the split mode:

Split mode	What counts as one shot
By paragraph	A blank line starts a new paragraph. One paragraph per shot
By line	Each non-empty line is one shot
By sentence	Split at periods, question marks, and exclamation marks. One sentence per shot

For example, with sentence splitting, a script with 6 sentences produces 6 short segments joined into 1 video. Empty segments are dropped.

Static templates read and lay out the text according to the split result. With image or video templates, the left side of the input area is the voiceover; click ✨ to generate a visual prompt per shot. If the toggle is on and the right side already has a visual prompt, compositing uses that text directly and the backend does not ask the LLM to rewrite it. If the toggle is on but the right side is still empty, the LLM fills in a visual prompt from the voiceover. If the toggle is off, the visual description uses the voiceover text as-is.

Fixed Script does not require an LLM when using a static template or when the visual prompts are already written.

Batch Create

Use this when you want many videos with different topics using the same template and voice. One topic per line in the text box, up to 100 lines. Blank lines do not count.

One topic line corresponds to one complete video and is also that video's title. It is not one shot. The shot count still applies: it is the number of shots inside each video, shared by all topics.

For example, with 3 lines and a shot count of 4, 3 jobs are queued. Each job is a complete AI Create run: write 4 voiceover lines for that topic, produce 4 picture segments, and join them into 1 video. Template, voice, character, subtitles, and background music are all shared. A failed job does not affect the others; view or cancel progress in the Job Center. Batch Create requires an LLM and does not have Pause after storyboard generation.

Picture, Sound, and Subtitles

First choose the aspect ratio (portrait 9:16, square 1:1, or landscape 16:9), then the template mode, then the specific template:

• Static style: laid out locally; does not call image or video generation and does not ask the LLM for visual prompts. The voiceover text appears on the picture and is read aloud.
• Generate images: each shot uses a text-to-image (t2i) workflow to produce one image, which is then joined with that line's voice into a short segment.
• Generate video: each shot uses a video workflow to produce a short segment. By default only text-to-video (t2v) packs are listed. After picking a character, image-to-video (i2v) packs are added: a normal i2v pack sends one camera-view image, an i2v_ref pack can take multiple reference images (Quick Create still sends one image for the current camera view), orbit does a 360° orbit, and extender is a long-video segmentation pack mainly used by the Director and not recommended for manual selection in Quick Create. If the pack declares acceleration, Turbo can be enabled. Packs whose names contain enhancer enhance the prompt inside the pack.

Voice, subtitles, and background music are below the template and shared by all three script modes:

• Static and image templates only support custom voice: synthesize each voiceover line and replace the audio track of that short segment.
• Generate Video can also use model audio: keep the audio produced by the text-to-video model instead of replacing it with a voice. Subtitles are still written from the voiceover text and that segment's duration.
• When the voice is local TTS (text-to-speech) or API TTS, you can preview it under the voice, see Text-to-Speech. If the voice provider is ComfyUI, choose a speech workflow separately; the voiceover text is filled in automatically.
• Subtitles can be turned off, and their font size and color can be changed. The title and the on-screen voiceover text can also be hidden or scaled separately.
• Background music is chosen from the Music Library and mixed in after the whole video is joined. A volume of 0.1 to 0.3 is recommended.

Characters and Camera View

The character is optional. All three script modes can pick one, and Batch Create shares the same character across all topics.

After selecting one, two different things happen.

The first always happens: the character's appearance description is prepended to each shot's visual prompt, so the shots are textually the same person. Layout sentences such as "16:9 landscape, four views" in the character sheet description are removed first, to avoid the picture being drawn as a four-panel collage. The style prefix you enter on the page is placed after this appearance text.

The second happens only when the video pack actually has an image input slot. Pure text-to-video packs have no image slot, so the character image is not uploaded. In that case the camera view is only a text hint to the LLM, for example that the current reference view is front, so it can write the visual description from that angle.

After selecting a character, you can switch the video workflow to an image-to-video pack. Such packs send one character image per shot according to the view, to lock the same person:

• By default the view you selected is used: facial half-body close-up, full-body front, full-body side, or full-body back.
• When the LLM writes the storyboard, it can switch a particular shot to another view, which uses multiple character sheet images.
• Choosing Character sheet sends the entire four-view image for every shot instead of splitting it into individual views.
• If that view has no image yet while the workflow requires a reference image, the job fails directly.

Image templates only write the appearance into the prompt and do not send the character image into the text-to-image pack.

Which Text-to-Video (t2v) Pack?

Pack	Text side	Notes
minimax_h3_t2v	Official TE	Pure prompt-to-video; Turbo can be enabled
minimax_h3_t2v_uncensored	Uncensored NVFP4 TE	Uncensored, no enhancement (corresponds to the local canvas Text-to-Video); LightX2V 8-step Turbo by default; Turbo can be enabled
minimax_h3_t2v_enhancer	Uncensored INT8 TE + tail	Uncensored + prompt enhancement; requires Guide; the enhanced text is visible in the result; Turbo can be enabled

None of the three accepts a first frame. To give each shot a character image, switch to minimax_h3_i2v / minimax_h3_i2v_uncensored / minimax_h3_i2v_enhancer.

Character Library

The sidebar's Resources → Character Library stores a character's name, appearance description, and one 16:9 four-panel character sheet. The sheet has four fixed columns from left to right:

From left to right	Content	Stored as	Used by
Column 1	Full-body front	Full-body front	Image-to-Video camera view, Motion Transfer <Picture 1>
Column 2	Full-body side	Full-body side	Motion Transfer <Picture 2>; fills multi-reference packs
Column 3	Full-body back	Full-body back	Motion Transfer <Picture 3>; fills multi-reference packs
Column 4	Oversized right-side facial half-body close-up	Facial close-up	Digital Human, Image-to-Video close shots

The full sheet is only for preview and splitting; it is never used as a video frame. Choose one add method on the page; the other method's button does not appear at the same time.

Generate from Description

Use this when you have no image and only an appearance. Enter a name, plus age, face shape, hairstyle, and clothing. You can also upload 1 to 3 reference images (or choose them from the library); when reference images are present, the appearance description can be left empty. One image uses the single-reference workflow in System Settings; 2 to 3 images use the multi-reference workflow. The first image is first pasted onto a 16:9 light-gray canvas to guarantee a landscape character sheet. You can click ✨ to use the default LLM in System Settings with the prompt template Character Library · Four-Panel Character Sheet (character_sheet). Optimization adds matching face shape, headwear, bottoms, shoes, and bags, and writes the fixed four-panel composition and quality requirements.

Click Generate Character Sheet. This creates the record and queues image generation itself; you do not need to click Save afterward. The image provider, character sheet workflow pack, single-reference workflow, and multi-reference workflow are all in System Settings → Default Character Generation, not on the character page. With no reference images the default pack is qwen_image_21 at 16:9 (1536×864; if the selected pack's width/height limit is smaller, it shrinks automatically). If you did not click optimize, the backend still appends the four-panel composition after the appearance description. If you did optimize and the prompt already contains "16:9" or "full-body front", it is not appended again.

Upload a Character Sheet

Use this when you already have a finished landscape four-panel image. You can also choose one from the library next to the upload box. From left to right it must be full-body front, full-body side, full-body back, and right-side facial close-up. Use a light or solid background and leave vertical gaps between the four figures. The appearance description can be left empty; if filled in, Quick Create prepends that text to every picture.

Click Save and Auto-Split. It splits and creates the record immediately without calling the image provider.

Do not put these two kinds of image into this upload box:

• A three-panel front/side/back turnaround. The fourth column is treated as the facial close-up, and the back may be cut into the side.
• A single front-facing portrait. Without vertical gaps it is split into four proportional strips, and the cuts may land on the person.

Split rules: find the gaps between the three figures from the brightness changes in each column, and cut at the midpoint of each gap. If there are fewer than three gaps, fall back to roughly 23% / 22% / 23% / 32%. If the right-side close-up still carries a gray edge from the neighboring column, move the cut right along the top color-block boundary. Click a card thumbnail to open the large view and switch left/right among the facial close-up, full-body front, full-body side, full-body back, and the original sheet.

Image files are under backend/data/characters/.

Storyboard Templates

Settings → Storyboard Templates. Upload/delete storyboard templates. Filename prefixes: static_ / image_ / video_. Canvas size must match the aspect-ratio directory. Built-in entries cannot be deleted; user files with the same name override built-in entries.

ComfyUI Model Directory (Quick Reference)

Purpose	Directory (relative to models/)	Common files
Image/video main model (UNET)	diffusion_models/	qwen_image_2.1* / qwen-image-2.1* (do not mix minimax_h3_*)
Text encoder / generation tail	text_encoders/	Images use qwen3vl_8b*; video H3 uses qwen3vl_32b* (do not mix)
VAE	vae/	Images use qwen_image_2.1_vae*; video uses minimax_h3_*_vae*
Turbo / UC / style LoRA	loras/	Video Turbo, image uncensored, PoseStudio, outfit swap, color grading, and their respective LoRAs
Legacy SD checkpoint	checkpoints/	Only t2i_basic
Upscale model	upscale_models/	RealESRGAN used by upscale_basic
SeedVR2	SEEDVR2/	seedvr2_ema_3b_fp16.safetensors + ema_vae_fp16.safetensors (numz nodes)
ControlNet	controlnet/	Qwen-Image 2.1 Fun ControlNet Union

---

Related: Image / Speech / Tools; Settings → System Settings.