MiniMax H3 TF2VA · Nova.4 — Text and First-Frame Video with Optional Audio
Workflow version: Nova.4
Guide updated: 17 November 2026
Generate video from text, a single starting image, or several images used for separate clips. An optional reference video supplies timing, cuts and scene descriptions. Use H3-generated sound, keep the reference video's audio, or attach a chosen excerpt from a dedicated recording.
This is the video-generation edition of the Unified workflow. It keeps prompt review, scene analysis, saved runs, per-part corrections and optional upscaling. Loaded audio is a soundtrack: it does not drive lip synchronization or set the video duration. The music-generation and lip-sync branches from the combined edition are not part of this workflow.
New in Nova.4: separate opening and later-render warm-ups let each scene's prompt establish its camera angle and pose before the visible clip begins. The workflow generates a hidden lead-in, discards it from video and generated audio, and keeps the planned clip duration and cut times. Both warm-ups are set to 1 second; set either to 0 to disable it.
TF2VA means text / first-frame to video and audio here. The current planner assigns one starting image to each clip. It does not supply a separate ending image, even if an inherited reference_mode dropdown still displays first_and_last. Two loaded images normally create two clips; they are not the endpoints of one clip. First/last interpolation, FMF and Ref2VA multi-reference modes from the combined guide do not apply to this edition's planner.
First test: use one image or text alone, a five-second duration, manual prompts, both warm-ups at one second, and one render part per queue. Leave reference-video analysis, dedicated audio and upscaling off until the base result works. Enable Disk + Keep in output so the result can be continued or corrected.
What this version supports
Text-only clips: the first prompt creates the opening clip; each nonblank next-prompt line adds another clip.
One image, several shots: reuse the same image for different scene descriptions, actions or camera angles in the first/next prompt list.
Several images: create one clip per enabled image, in loader order, with a separate prompt for each. Images are not described together as one scene.
Optional LLM descriptions: expand each clip's instructions using its assigned image, or text alone. A single image can be described without loading a video or running scene detection.
Reference-video analysis: measure cuts and describe the action between them, then combine those directions with the assigned image when one is loaded.
Final prompt review: inspect the complete generation prompt for every planned scene, including LLM results, and edit it before rendering.
Hidden warm-up: give each scene's prompt time to establish camera and pose, then discard the lead-in while keeping the planned visible duration.
Selected soundtrack excerpts: choose a recording's Start and Duration. Loop or trim that excerpt to the independently planned video length.
Saved-run recovery: continue an unfinished run, regenerate an ending, update one saved part, or load an existing result for export and upscaling.
Install and select your models
Use ComfyUI with native MiniMax H3 support. The current workflow checks use ComfyUI 0.42.1. Update the backend if its H3 conditioning or audio-decoding nodes are missing, then refresh the browser after restarting.
Upgrading to Nova.4: update Eclipse Nova to 1.0.44 or newer, restart ComfyUI, refresh the browser and open the Nova.4 workflow. It includes the new conditioning and append nodes required for warm-up cropping. Select New run to apply the new settings; existing saved runs retain their original render plan.
Install or update these packs and their dependencies in ComfyUI's Python environment:
ComfyUI Eclipse Nova — source controls, scene analysis, planning, prompt review, render loops, saved runs, upscaling and saving.
ComfyUI Smart Model Loader Nova — the H3 loader, model/LoRA presets and Download Manager.
ComfyUI SmartLLM Nova — optional image descriptions, text expansion and reference-video descriptions.
ComfyUI-sol-attn — the included H3 optimization and Diff-Aid patches.
Block Sparse Attention is supplied by ComfyUI. The optional Upscale Video node is supplied by Eclipse Nova. Eclipse also supplies the loops; Easy-Use is not required. SmartLLM's nodes must be installed to load the complete graph, but its model does not run when both analysis and LLM descriptions are disabled.
H3 generation models
Select an FL2VA-compatible MiniMax H3 model, its H3 text encoder, video VAE and audio VAE. The Comfy-Org MiniMax-H3 repository provides the official files.
Diffusion model: for example,
minimax_h3_fl2va_pruned_int8_convrot.safetensors. Place it inmodels/diffusion_models/.H3 text encoder: for example,
qwen3vl_32b_minimax_h3_int8_convrot.safetensors. Place it inmodels/text_encoders/.Video VAE: for example,
minimax_h3_video_vae_fp16.safetensors. Place it inmodels/vae/.Audio VAE: for example,
minimax_h3_audio_vae_fp32.safetensors. Place it inmodels/vae/.
The saved workflow can contain a community checkpoint or a locally renamed encoder. Select the filenames actually installed on your computer. The audio VAE is still needed for H3's joint generation path, including when an external recording will replace the generated soundtrack.
If your preset list contains MiniMaxH3_TFL2VA, that older preset name can still be used for the matching native H3 model configuration. Check both Smart Model Loader and the separate LoRA Stack; changing one preset does not automatically change the other. The TF2VA workflow name does not rename your local presets.
Check every enabled LoRA and choose sampling settings that match it. Disabled rows do not require their files. The currently inspected loader uses Euler / Simple, 8 steps, CFG 1, video shift 12 and audio shift 5.5; these are saved settings, not a universal recommendation for every checkpoint. Use the selected model or acceleration LoRA's supported settings.
Downloads and the description model
For a missing resource with a saved source reference, open its Source / integrity controls, or the LoRA row's i button, and choose Open in Download Manager. Review the intended file and destination, add it to the queue, then explicitly start the download. Keep the workflow open for automatic assignment; after a browser reload, select the downloaded filename yourself. See the Download Manager guide.
Smart LM uses a separate model for descriptions. The saved example selects a Qwen3-VL 8B vision model; another supported vision model or backend can be used. H3's 32B generation encoder does not replace that description model. Text-only expansion can use a supported text model, but image/video descriptions require vision support.
Choose resolution, render length and memory settings for your hardware. Start with a short clip and keep upscaling off while checking model loading and decoding; an older guide's 16 GB configuration is not a requirement or a capacity guarantee for this version.
Find the controls
START HERE · optional inputs: Toggle Images, Audio override, Reference video, Live analysis and LLM descriptions.
Shared H3 loader + separate LoRA presets: Select models, LoRAs, sampling and memory settings.
Canvas + output folder · 24 FPS: Set canvas, output folder and Smart Folder's Duration.
Optional images · one loader + resize: Load, enable, order and crop the starting images.
Reference video · timing + fallback audio: Load source clips, trim them and choose whether to keep their audio.
Optional dedicated audio · excerpt selection: Select a soundtrack and its Start / Duration.
Global direction · every generated clip: Enter instructions shared by every clip.
Manual cut times + first / next part prompts: Enter each clip's direction and optional cut times.
MMH3 Reference Analysis · optional: Detect source-video scenes and describe sampled frames.
Native clip plan · images / timing / sound: Set opening and later-render warm-ups, resolve clip timing and image assignment, and inspect H3 Plan Report.
LLM descriptions · optional · image or text: Refine each planned scene's complete instructions.
Part source · final generation prompts: Inspect and edit the final text that reaches generation.
Run operation + saved-run selection: Choose New, Resume, Update or Load; select storage and saved run.
Native H3 render loop · saved inputs + resume: Set the render seed, part limit and restart part; view the accumulated video.
Optional upscale · enable group to use: Upscale the assembled result before saving.
Shared result + save video: Save the selected assembled result with its soundtrack.
Use Smart Folder's Duration for the clip-duration rules below, including 0 to preserve loaded reference-video timing. Its value is connected to the planner, so editing a downstream connected duration widget does not change the source value. Keep H3 at 24 FPS and use a canvas divisible by 32.
First run: one image or text alone
Load the matching H3 model and LoRA settings. Select New run in Video Run Storage.
Turn Reference video, Live analysis, Audio override and LLM descriptions off. Leave upscaling muted.
For image guidance, turn Images on and select one enabled entry in Load Image Timeline. Check its crop against the chosen canvas. For text-only generation, turn Images off.
Set Smart Folder's Duration to
5. Clear manual cut times and the next-prompt list.In Clip timing + prompts + hidden warmup, keep opening_warmup_seconds = 1 and warmup_seconds = 1. These hidden seconds do not reduce the five-second output.
Write the shared style in Global instructions · every clip and the opening action in First clip · LLM instructions / direct prompt.
Enable Disk + Keep in output. In Video loop · saved inputs + resume, set Parts per queue = 1.
Enable Pause on Final generation prompts if you want to inspect and edit before sampling. Queue, review the selected prompt, and press Continue when it pauses.
Inspect the assembled video. If the plan contains more render parts, select the saved run and use Resume saved run to continue it.
With no external audio source, H3 generates the sound. Include ambience, effects or spoken dialogue in the prompt when wanted. A loaded soundtrack replaces that generated audio during assembly; it does not synchronize the generated performance to the recording.
Describe a single image without scene detection
Keep Images on, Reference video off and Live analysis off, then enable LLM descriptions. Choose an image-capable task such as MiniMax H3 Scene or a suitable description task in that group's Smart LM Loader.
Write what the clip should do in the first prompt. The LLM receives the image together with the global and first-clip instructions. Its result appears in Final generation prompts before H3 uses it. There is no video scene-detection pass in this setup.
To use the same image for several angles or actions, add one nonblank line per later clip to the next-prompt list. Each clip is described separately using the same image and its matching instruction. With Images off, the same flow works from text alone; choose a task that accepts text input.
How images and prompts create clips
The following rules apply when Reference video is off:
No images: an opening clip plus one per nonblank next-prompt line. Each clip is text-only, using the global direction and its matching first/next instruction.
One image: an opening clip plus one per nonblank next-prompt line. Reuse the single image for every clip.
Two or more images: one clip per enabled image. Image 1 gets the first prompt; later images get the corresponding next prompt.
A blank first prompt still reserves the opening clip. With no next prompts, no images and no video, there is one text-only clip using the global direction.
For several images, fewer descriptions are allowed: an image without its own later instruction uses only the global direction. It does not inherit the previous clip's action. More prompts than images produce an error, because this mode keeps one clip per image.
Keep images in the intended loader order and disable unused entries. The optional description stage processes them individually. Keep its Together off / individual images setting; combining those images into one scene would lose the one-prompt-per-clip pairing.
The assigned image anchors the start of each generation. Independent clips can differ in pose or motion, even when they reuse the same image. Image-timeline fades do not create crossfades between the finished H3 clips.
Global, first and next prompts
Global instructions: style, shared subject constraints, overall direction and any generated-sound instructions that apply throughout.
First clip: the opening framing, starting pose, action and camera movement.
Next clips: one nonblank line per later clip. Each line is a separate instruction, not another paragraph for the first clip.
Manual cut times: optional comma-separated boundaries in seconds. These control timing; mentioning a cut in prose alone does not create a measured planner boundary.
For example, use one image or no images with:
Global: Cinematic realism, warm evening light and restrained camera movement. Soft wind and distant city ambience.
First: Medium shot. The subject stands beside a rooftop railing and slowly turns toward the skyline.
Next, line 1: Close side profile. The camera makes a gentle push-in as the subject watches the sunset.
Next, line 2: Wide rear view. The subject rests both hands on the railing while the city lights brighten.
This creates three clips. With three images, the same prompt list assigns one description to each image instead.
Duration and manual cuts
Video duration is independent of the selected soundtrack's length.
No reference video, cut times blank: each logical clip lasts Smart Folder's Duration.
No reference video, explicit cuts: cuts divide the same total of clip count × Duration.
Reference video, Duration
0: preserve each loaded clip's trimmed duration.Reference video, Duration greater than
0: override each loaded source clip with that duration; scale its scene boundaries with it.
Preserve source timing: enable Reference video and set Smart Folder → Duration = 0. The existing duration connection passes zero to the planner; no rewiring is needed. Restart ComfyUI and refresh the browser after updating Eclipse Nova to load the expanded duration range. Use a positive duration when returning to image/text-only clips.
Without video, Duration must be positive. For N clips, supply exactly N−1 cut times, or leave the field empty. Times must be increasing, greater than zero, and before the total video end. H3 maps them to its 24 FPS timeline, so extremely close boundaries may collide.
Example: three prompts with one image, or three separate images, at 4 seconds per clip:
Manual cut times blank: clip lengths are 4, 4, 4 seconds; total duration is 12 seconds.
Manual cut times
3, 8: clip lengths are 3, 5, 4 seconds; total duration is 12 seconds.
With only the first prompt and one or no image, there is one clip. To create additional manually timed clips, add matching next prompts as well as their cut times. Clear stale example cuts when returning to one clip.
Hidden warm-up for camera and pose
In Native clip plan → Clip timing + prompts + hidden warmup, use:
opening_warmup_seconds: hidden lead-in for the first render.
warmup_seconds: hidden lead-in for every later independent render, including length-based splits of a long scene.
Nova.4 sets both to 1 second. Set either to 0 to disable it. For example, set only opening_warmup_seconds = 0 to retain the reference-image opening while later clips still receive warm-up. Nodes added to an existing graph default to zero, preserving earlier behavior.
The new scene's complete reviewed prompt applies during both the hidden lead-in and visible footage. When an image is assigned, generation starts from that image and has time to move toward the requested framing and pose before the intro is discarded. This is useful when one image supplies several shots, and also works with separate images or text alone. One second gives the scene time to settle; it does not guarantee that every requested change has finished by then.
Both settings accept 0–10 seconds, rounded to the 24 FPS frame grid. The selected warm-up must leave room for visible frames inside Max clip frames. Review the output and increase the relevant warm-up if the reference-to-scene transition is still visible. Reduce it if the hidden intro removes an opening action or generated dialogue you wanted to keep.
For example, a 5-second clip with 1 second of warm-up keeps five visible seconds after discarding the first 24 frames. Its cut times and selected soundtrack stay on the original output timeline. Generated audio is cropped with the video.
The workflow uses MiniMax H3 Clip Conditioning_v2 and Append Video Clip_v2, with both crop_start and keep_frames connected. Restart ComfyUI and reopen the updated workflow to load these nodes. Use New run to apply changed warm-up settings; Resume and Update retain the saved settings and render layout.
Logical clips versus render parts
Max clip frames limits an individual H3 render. The saved setting of 362 is about 15.08 seconds at 24 FPS. Longer scenes split into several render parts that keep the same prompt and starting-image assignment.
Warm-up frames count toward that limit. With a one-second lead-in, a 362-frame render has room for up to 338 visible frames, about 14.08 seconds. Warm-up can therefore increase the number of render parts even though the final duration stays the same.
H3 renders supported lengths on its 17k + 5 frame grid, with at least 124 frames per invocation. Padding is discarded to keep the planned output length. A two-second scene can therefore cost substantially more than two seconds' worth of sampling; increasing the number of cuts can make generation much slower.
These are independent renders, without guaranteed motion or generated-audio continuity at their joins. Use H3 Plan Report and the saved part ranges to distinguish source clips, detected scenes and actual render parts. Restart part refers to render parts, not necessarily the scene number shown in prompt review.
Reference video: timing, cuts and descriptions
Enable Reference video, load one or more clips and select the wanted source ranges. Keep playlist transitions set to Cut. Set Smart Folder → Duration = 0 to follow the loaded timing. A positive duration instead overrides each source clip.
With Live analysis on, the workflow detects cuts and describes the sampled frames in each scene. Its cut times and first/next scene descriptions replace the manual cut/first/next fields. Global instructions still applies. With Live analysis off, the manual fields are used while the loaded video still supplies its timing and optional audio.
The video is analyzed for text and timing; its frames are not passed to H3 as a video-to-video conditioning stream. When images are enabled, their assigned starting images supply the visual guidance. Without images, H3 generates from the resulting descriptions.
Image assignment with a video playlist
With reference video active, images are assigned by loaded source clip, not by detected scene:
One image is reused for all source clips and their scenes.
Several images cycle in loader order: source clip 1 uses image 1, source clip 2 uses image 2, and so on, returning to image 1 when needed.
Every detected scene and technical render split inside a source clip keeps that source clip's image.
Images beyond the number of loaded source clips are unused. For one video file, loading three images does not create three image clips; only image 1 is assigned.
Each loaded source-clip boundary is preserved even if detection or cut limiting would otherwise merge similar-looking clips. Additional edits inside those files can create more scenes. With Live analysis off and blank cuts, a prompt list longer than the video playlist adds evenly spaced prompt boundaries across the loaded timeline while retaining the source boundaries. Check the plan report when mixing source clips and manual descriptions.
Choose the task and appearance policy
The video-analysis Smart LM task and Scene Analysis to H3 → appearance mode are manual choices. Choose a matching pair:
Transfer framing and movement while the target image supplies appearance: choose the Scene Motion Analysis task and Exclude source appearance.
Describe the source fully for generation from text: choose the Scene Description (T2V) task and Include source appearance.
Use an ordinary image description or image-capable H3 task: choose, for example, Detailed Description or MiniMax H3 Scene, with Include source appearance and Scene JSON on.
Scene Description (T2V) still receives visual samples. Its name refers to generation from the description. The separate MiniMax H3 T2VA Timeline task is text-only and does not analyze connected media.
Changing the appearance dropdown does not rewrite a full caption into motion-only text. Change the task too and rerun the analysis. The dedicated scene tasks already produce the collector's structured observations. For ordinary caption/H3 tasks, Scene JSON packages their answer for Scene Description Collector; leave it off in the separate LLM descriptions group, where ordinary text goes directly to prompt review.
Detection and sample coverage
Scene Detector scans every source frame on CPU. The default longest side is 128, with adaptive threshold 2.5, window width 4, minimum content 15 and minimum scene length 1. The short-blend check uses gradual_threshold = 2.3 and gradual_min_change = 0.2.
The updated blend check rejects steady colour drift from smooth zooms and pans instead of treating that alone as an edit. It still detects concentrated appearance transitions. This is a heuristic: abrupt reframing can resemble a cut, and subtle or long transitions can be missed. Restart after updating and rerun live analysis to replace old detected boundaries.
Scene Cut Limit → 0 means unlimited, not “disable cuts.” A positive value limits retained cuts per second: 1 requires at least one second between them; 2 requires half a second. Source-file boundaries are restored after this limiter. Turn Live analysis off to bypass detection and enter a complete manual plan; a multi-file source playlist still retains its file boundaries.
Scene Analysis Plan selects the samples sent to the LLM. The inspected setup uses 16 samples per scene, 16 per window and 2 overlapping samples when additional windows are needed. Larger sample counts can reveal brief actions, but they do not add measured cuts.
For video scenes, use Frames + Together on or Grid so the model can compare ordered frames. The saved Grid example uses eight frames per grid and 512-pixel tiles. Keep the LM stride at 1 when the plan already provides a reduced selection. Grid/window splitting does not create extra output scenes.
Keep Smart LM Unload connected before H3 generation when the two models share GPU memory. Thinking is optional; leave it off for an initial description pass. If a supported backend spends too long reasoning, disable Thinking or adjust its output budget. The ■ Stop LM control interrupts a stuck description run through the backend's supported stop behavior.
Optional LLM descriptions and final prompt flow
There are two optional analysis steps with different inputs:
Live analysis observes the reference video's sampled scenes and produces their timings and descriptions.
LLM descriptions receives each planned scene's complete instructions plus its assigned starting image, if available. It can refine those directions without any reference video.
They can be used independently or together. The flow for a New run is:
Manual first/next directions OR live video descriptions
+ Global instructions
↓
Clip timing + image assignment
↓
Optional LLM description for each planned scene
↓
Final generation prompts
↓
Reviewed plan → H3 render loop
With LLM descriptions off, the global and matching scene instructions go directly to final review. With it on, the LLM result replaces that complete input prompt. The raw global/first/next text is not appended a second time. Important shared instructions must therefore be present in the LLM's final answer; check them in the review.
The connected prompt list takes priority over Smart LM Loader's local Prompt field. Use the workflow's Global and First/Next fields for normal guidance. Enable Merge user prompt only when you deliberately want the loader's local text added to every request. This prevents an apparently filled local field from being silently mistaken for the active input.
The optional group has its own LLM descriptions · one per clip review. Part source → Final generation prompts is the final checkpoint: it shows the resolved text actually applied to the plan, including any upstream description edits. It displays prompts after the relevant nodes execute; an empty editor before the first queue is not evidence that generation will receive an empty prompt.
Review and edit before rendering
Enable Pause on Final generation prompts before queueing New run.
When the prompts arrive, use the dropdown or ◀ / ▶ buttons to select each logical scene, including the opening entry.
Edit its full generation prompt. Keep the intended global style and sound instructions, because the edited text replaces that scene's complete prompt.
Press Continue to use all edits in the waiting run. The upstream LLM does not need to run again merely to apply them.
With Pause off, the first execution displays and passes through the incoming text. Displayed prompts remain editable: edit and queue again to apply those changes on the next execution. Turning Pause off after Continue retains the reviewed text for matching inputs. Changed upstream prompts or ordering start a fresh review instead of applying unrelated drafts.
Reset this prompt restores its incoming text. Cancel run cancels a waiting review. Do not delete entries to change clip count or timing; change the source fields and start a New run instead. Technical render splits share their scene's reviewed prompt.
Prompt display does not require saving a separate prompt file. Saved generation runs retain the resolved prompts for recovery. For a correction to an already saved render part, use Part Prompt Editor, as described below.
Dedicated audio and excerpt selection
Soundtrack priority is:
Audio override: the selected excerpt from the dedicated Load Audio node.
Reference-video audio: when no override is supplied and the loaded sources have usable, unmuted audio.
H3-generated audio: when neither source provides a recording.
To use part of a longer recording:
Turn Audio override on.
In Dedicated audio · select excerpt, choose Selected file and your recording.
Set Start to the excerpt's beginning in seconds.
Set Duration to the amount to keep. Audio Duration
0keeps the remainder from Start.Set video duration separately in Smart Folder and inspect the plan report.
For example, Start = 35 and Duration = 12 selects seconds 35–47. A 20-second output repeats only that selected 12-second excerpt as needed. An eight-second output uses its first eight seconds. Material before or after the selected excerpt is not brought into the loop.
The final assembly trims a long excerpt or repeats a short one to fit. It does not time-stretch the recording, detect its beat or synchronize lips to it. Align a dedicated soundtrack with reference-video action yourself; source cuts are measured from the loaded video's start, not from the full recording's absolute time.
Mute the source-video slots and disable Audio override to use generated sound. A present but silent recording still counts as a selected soundtrack. Resume and Load restore their saved audio selection, so changing the current audio controls does not retrim an existing run.
Save, continue and correct a run
In Video Run Storage, choose the operation and storage mode:
New run: Build a fresh plan from the current sources and prompts.
Resume saved run: Continue unfinished parts, or regenerate from Restart part N onward into a new revision.
Update selected run: Replace exactly one saved part N, keeping the other parts and backing up the previous version in the same folder.
Load saved run: Load the current run or a selected backup for preview, upscale or export, without generation.
Disk + Keep in output retains completed parts and saved inputs across ComfyUI restarts. The saved output directory runs places them under ComfyUI/output/runs/. The final delivery video's Smart Folder destination is separate. Disk + Temporary uses cache-lifetime storage; RAM keeps decoded parts in CPU memory and is not persistent across restarts. Preview and export still write video files in RAM mode.
A saved run includes its timing, source/image assignment, resolved prompts and selected soundtrack. Resume restores those inputs without fresh analysis. Model objects are not stored: choose compatible model, LoRA and sampler settings when continuing.
Continue or regenerate an ending
Click Refresh runs in Load Video Run and select the result you want to continue.
Choose Resume saved run and keep Disk + Keep in output enabled.
Use Restart part = 0 to continue after the last completed part, or a one-based part number to regenerate that part and everything after it.
Set Parts per queue:
1for a single review step,0for all remaining parts, or a larger limit for a batch.Change Render seed · change before queueing when you want a new result, then queue.
Select the newly created revision before continuing it again. Set Restart part back to
0after a replacement when you want normal continuation.
Queueing New run again starts another fresh run. A completed run resumed with Restart part 0 has nothing left to generate.
Stop after current part lets the current render finish, saves it, updates the accumulated preview and then stops the loop. An immediate ComfyUI interruption can discard the unfinished part; earlier completed parts remain available. Changing the seed while a queue is running does not change that queue's remaining work.
Update one saved part and its prompt
Select the run with backup = Current run.
Choose Update selected run, keep Disk + Keep in output, and set Restart part to the saved part's one-based number.
To change its text, click Read selected run in Part Prompt Editor, select that render part, edit the complete combined prompt and enable Apply edits. Clear pending changes to other retained parts.
Optionally change Render seed and queue. Update renders exactly that part, regardless of Parts per queue.
Other saved parts remain intact. Each successful update backs up the previous version, including its prompts and media. Select a backup in Load Video Run to preview it with Load saved run, or use Restore selected backup to make it current without rendering. Restore also preserves the version it replaces.
For several changed parts or a new ending, use Resume from the earliest changed completed part. Ordinary manual prompt boxes and Final generation prompts belong to New run; they do not replace the stored prompts on Resume or Update. Use New run for changed timing, source images, video selection or audio excerpts.
Keep the whole saved folder when moving or backing up a run, including its hidden history folders. Do not delete individual part files to force regeneration. To discover a moved run, place its complete folder under ComfyUI output and refresh the run list.
Optional patches and upscaling
H3 model patches
The model order is LoRA → H3 patches → Sparse → Diff-Aid → sampler.
Choose Scheduled Sol Attention or Native Sparse Attention. Leave Scheduled Sol off while Native Sparse is enabled.
Fused Modulation is a separate optimization and can be used with either attention choice.
Chunk Feed-Forward can reduce peak memory at a speed cost.
Diff-Aid Text Modulation adjusts text conditioning and runs last; it is not a speed patch.
Use each group's Patch switches · OFF = bypass bar. For a patch with its own enabled setting, leave that setting on when controlling it through the switch. Mute a whole group to omit it; named routing falls back to the preceding available model. Keep prompt, seed and sampling fixed when comparing patches.
Upscale the assembled result
Enable the whole Optional upscale · enable group to use group and select an installed model in Upscale Video. PyTorch: models come from models/upscale_models; TensorRT: engines come from models/tensorrt/upscaler and require a compatible runtime, GPU and engine profile.
Choose the model's native scale, a scale such as 2×, a long-edge size such as 1080p/2K/4K, or custom dimensions. Start with batch size 1 if memory is limited. Upscaling preserves frame count, timing and soundtrack; it does not add motion frames.
It runs after the parts are assembled. The render-loop preview and saved generation parts keep their original size; inspect the final output to judge the upscale. Mute the group to restore the original result without rewiring.
For faster iteration, generate and review the base video first. Then choose Load saved run, enable upscaling and export once. Repeatedly upscaling a growing partial run repeats work on its accumulated frames.
Final video export
Shared result + save video prefers the active upscaled video, then the original assembly. The selected recording or generated sound is already carried by the assembled VIDEO.
Choose the output folder in Smart Folder and check Save Video Data's filename and delivery settings. The current profile uses H.264 MP4, CRF 19 and Veryfast, with workflow and generation metadata enabled. Optional trimming is a separate final-export operation; leave it disabled when checking the planner's complete duration.
The final MP4 is the delivery file. Keep the saved run folder as well when you want to resume, update parts or restore earlier versions.
Common fixes
Two images do not interpolate into one clip: this planner assigns one starting image per clip. Its inherited
first_and_lastcontrol does not add an ending image. Use the TF2VA clip rules above.Text-only generation still uses an image: turn Images off in START HERE. The planner's assigned image takes priority over the inherited reference-mode dropdown.
A single image only creates one clip: add nonblank next-prompt lines. The first prompt alone intentionally creates one clip.
Every shot visibly starts from the reference pose: enable warmup_seconds for later renders, and opening_warmup_seconds when the opening also needs settling. Check the new crop connection and start a New run. Increase the lead-in if the scene still needs more time to reach its requested framing or pose.
Extra images are ignored with video: assignment follows loaded source clips. One source video uses one image throughout its detected scenes.
The total duration is unexpectedly multiplied: without video, Duration is per clip; with video, a positive value overrides every loaded source clip. Set Smart Folder → Duration = 0 to preserve video timing.
The audio changes the sound but not the action or duration: this edition attaches it as a soundtrack. Choose video length and motion prompts separately.
The wrong part of a long recording plays: check the dedicated audio's Start and Duration, then use New run. Resume and Load keep the saved excerpt.
Manual prompts or cuts are ignored: turn Live analysis off to select the manual source. Global instructions applies to either source.
The LLM ignores its local Prompt field: the connected per-clip list overrides it unless Merge user prompt is enabled. Enter normal directions in Global and First/Next.
Final generation prompts looks empty: queue the New-run route so planning and any enabled descriptions execute. Enable Pause before queueing to stop there for inspection.
Prompt Review edits do not affect a saved run: use Part Prompt Editor with Resume or Update. Edits to displayed New-run prompts can be queued with Pause on or off, but do not replace saved-run inputs.
The collector expects scene JSON: ordinary caption/H3 tasks need Scene JSON on in the video-analysis loader. The standalone clip-description loader should return normal text.
Source appearance leaks into an image-guided scene: choose Scene Motion Analysis with Exclude source appearance and rerun. Review the final prompt for conflicting details.
Zooms create unwanted cuts: update Eclipse Nova and rerun live detection. The improved blend check filters smooth drift, but abrupt reframing may still need manual boundary review. Cut Limit
0preserves all detected cuts.One scene produces an empty next-prompt list: that is valid. Do not add a dummy cut or prompt.
More LM samples do not fix a missed cut: sampling improves descriptions, while detection supplies boundaries. Use a complete manual timing/prompt set with Live analysis off when exact cuts are needed.
Resume generates nothing: Restart part
0on a complete run is already finished. Select an unfinished run or a valid replacement part.The run is absent from the list: use Disk + Keep in output, preserve the whole folder under ComfyUI output and click Refresh runs.
Memory runs out: reduce canvas size or Max clip frames, use Disk for accumulated parts, and leave upscaling off. Shortening only the retained duration below H3's minimum render length does not remove that minimum generation cost.
Further documentation
Native H3 clips and prompt planning — begin with the unified prompt-flow sections; the later Original Clip Planner section documents a different contract.
Scene analysis — detection, sampling, appearance policy and scene export.
Prompt Review — editing with Pause on/off and same-run continuation.
Saved runs and storage — persistence, retained inputs and recovery.
Loops, updates and backups — part controls, replacing saved parts and restoring versions.
Video upscaling — model formats, output sizes and memory use.
Smart LM Loader guide — tasks, image handling, Grid, Scene JSON and backends.
