BFS Shot H3 Conditioning:
BFSShotH3ConditioningSimple prepares one planned shot for MiniMax H3 video generation. It combines the shot’s prompt and reference pictures with visual information from the source footage, giving the model guidance for the scene, subject, and motion. You can guide generation frame by frame, provide the source as reference-video context, or use both approaches. The node also supports optional soundtrack conditioning, source-subject blurring to discourage copying an existing identity, and an experimental native temporal mode. It is the streamlined version of the H3 shot-conditioning node; controls for a setting picture, mask guide, and separate first-frame guide are available in the Advanced node.
BFS Shot H3 Conditioning Input Parameters:
shot
Required BFS_SHOT. The planned shot supplies the prompt, reference pictures, source frames, generation dimensions and length, and other shot-specific settings used to build conditioning. Connect the shot produced by your BFS shot workflow. There is no default value; the quality and completeness of the planned shot directly affect what this node can prepare.
clip
Required CLIP. This is the text encoder used to turn the prepared prompt into conditioning the video model can use. In the aligned-guide-plus-CLIP-video-context mode, it also provides the path through which Qwen sees sampled source frames. There is no default value; connect the encoder appropriate for your H3 workflow.
vae
Required VAE. The VAE encodes visual material into the latent representation used for generation and guidance. A missing or incompatible VAE prevents the node from preparing the shot correctly. There is no default value.
guide_mode
Required choice controlling how the source shot informs generation. Options are aligned guide (Add Guide) (default), native reference video (Video 1), both, none, aligned guide + CLIP video context, and ``TARA
- Native Temporal Alignment (experimental)
. An aligned guide places source-frame guidance alongside generated frames; a native reference video supplies the shot as video context;bothuses both paths. CLIP video context lets Qwen inspect sampled frames while motion still uses Add Guide, without native shot-reference latents. The experimental temporal mode uses a native motion-video path intended for its matching LoRA and has additional requirements described below. Choosenone`` when you do not want either of the standard source-video guide paths.
ref_image_size
Required choice controlling how reference pictures are sized for conditioning. Options are match (default) and max. Use match as the normal starting point when you want reference-image size matched to the shot; max requests the alternative maximum-size handling. Larger reference treatment may use more resources. This setting concerns reference pictures, not the generated video’s output dimensions.
audio_vae
Optional VAE for audio conditioning. Connect it when you want to use the shot’s soundtrack with with_audio enabled; without it, enabling with_audio does not attach that soundtrack in the shown implementation. No default connection is supplied.
with_audio
Optional boolean, default False. When enabled and audio_vae is connected, the node includes the shot’s audio in its conditioning path. This can help when matching mouth movement to speech. Leave it off when the soundtrack is unnecessary or you have no audio VAE.
vlm
Optional CLIP connection for a Qwen3-VL encoder. It can fill missing reference details left in the planner prompt, helping the final prompt describe supplied pictures. It is not required when your prompt already contains the details you need. No default connection is supplied.
inpaint
Optional choice, defaulting to the first entry in INPAINT_MODES. Its described modes are per-shot Planner behavior, forced mask-only generation, and off. Per-shot behavior follows each shot’s Planner mask mode; forced mask-only generation requests regeneration only where the shot mask permits; off disables this inpainting path. Mask-only work requires a usable person selection or mask on the shot. The supplied interface does not enumerate the complete INPAINT_MODES option list, so its exact displayed choices beyond the described modes are not specified here.
source_blur
Optional choice, default off. It changes what the model sees of the source subject in guide material, helping reduce the chance of copying that person’s identity while retaining useful motion or pose cues. Available choices are supplied by SOURCE_BLUR but are not fully enumerated in the provided interface; the tooltip describes region-and-detail combinations such as a masked area and specifies that masked area uses the shot’s own mask. Select a mode only when its chosen region can be identified in your footage.
source_blur_strength
Optional float controlling how strongly source blur removes detail. Default 1.0; minimum 0.1; maximum 1.0; step 0.05. The default is the coarsest setting described in the supplied context, leaving no fine detail in the blurred region. Lower values retain more source detail, which may preserve useful expression but may also make the old subject easier to copy. It matters when a blur-based source_blur mode is active.
source_blur_segmenter
Optional choice determining how regions such as a face, head, hands, person, or mouth are found for source_blur. Its default is the first entry in BLUR_SEGMENTERS; the complete option list is not supplied. The tooltip identifies SAM 3.1, which uses text prompts, and Sapiens2, which uses body-part classes and may download a roughly 1.6 GB model on first use. Modes that keep lips and tongue always use Sapiens2 for the lips. This selection affects which parts remain sharp or are altered.
source_blur_effect
Optional choice controlling the alteration applied by source_blur. Its default is the first entry in BLUR_EFFECTS; the tooltip describes blur and invert colours, but the complete option list is not supplied. Blur removes detail, while colour inversion retains shapes and expression without retaining the original colours or skin tone. Choose according to whether your shot needs expression cues or stronger removal of appearance detail.
clip_video_fps
Optional float used only with aligned guide + CLIP video context. Default 2.0; minimum 0.5; maximum 24.0; step 0.5. It sets how often Qwen samples the shot for visual context, not the generated video’s frame rate. Try 2, 4, or 8 fps: higher settings send more visual information but increase context use and encoding time.
model
Optional MODEL connection in ordinary use, but required for native temporal alignment. In that experimental mode, the model may be patched and its model output must be routed onward to the guider. No default connection is supplied. Connecting it in other modes also allows the model to pass through the conditioning workflow.
native_guide_scale
Optional float used only in native temporal mode. Default 1.0; minimum 0.125; maximum 1.0; step 0.125. It scales the motion video spatially before encoding without changing output size, frame count, or identity references. Try 0.5 when guide-token use or VRAM is a concern, recognizing that a smaller motion guide carries less spatial detail.
native_temporal_backend
Optional choice used only in native temporal mode. Options are flash partitions (no compile) (default), flex (original), and comfy kitchen (experimental). Flash partitions is described as preserving the temporal attention policy without compilation; Flex is the original implementation. The Comfy Kitchen option is explicitly experimental. This choice does not affect the standard aligned-guide or reference-video modes.
normals
Optional IMAGE containing MoGe normal maps of the shot frames. It is intended only for shots whose Colored Mark style is normal map. The source suggests producing it from BFS Shot Unpack frames through Run MoGe Inference and Render MoGe Geometry using normal_directx. Leave it unconnected for other styles; no default image is supplied.
BFS Shot H3 Conditioning Output Parameters:
positive
CONDITIONING containing the prepared positive guidance for this shot. It represents the prompt and applicable references or video-guidance context in a form the H3 generation workflow can use. Connect it to the downstream conditioning input rather than treating it as a finished image or video.
latent
LATENT prepared for generation of the planned shot. It carries the latent-side setup associated with the selected guidance and shot settings. Pass it to the downstream generation workflow; it is not a viewable frame sequence on its own.
prompt
STRING containing the prepared prompt returned for the shot. It may reflect reference-detail filling when a suitable vlm is connected and the planner left details to fill. Inspect it when you need to confirm the text being passed into conditioning. This is the third output socket.
model
MODEL containing the passed-through model or, in native temporal mode, the model prepared for that mode. Route this output to the guider when using native temporal alignment; otherwise, it can preserve a workflow that routes its model through this node. This is the fourth output socket.
BFS Shot H3 Conditioning Usage Tips:
- Start with
aligned guide (Add Guide)when you want the generated shot to follow source motion frame by frame. Trynative reference video (Video 1)when video context is more important than direct frame alignment, orbothwhen the workflow benefits from both forms of guidance. - For
aligned guide + CLIP video context, begin at the defaultclip_video_fpsof2.0. Raise it to 4 or 8 only when Qwen needs to observe faster or more detailed changes; this increases context and encoding cost without raising output fps. - When replacing a person, try
source_blurif the result keeps borrowing the source person’s appearance. Compare a blurred result withoffso you can judge whether identity separation improves without losing necessary pose or expression cues. - Enable
with_audioonly after connectingaudio_vaeand confirming the shot contains the soundtrack you intend to use. This is particularly relevant to speech and lip-sync work. - For mask-only generation, prepare a valid shot mask first and choose the appropriate
inpaintmode. Review the Planner’s per-shot setting before forcing the same mask behavior across every shot. - Treat
TARA - Native Temporal Alignment (experimental)as a separate workflow: connectmodel, route the node’smodeloutput to the guider, use 24 fps shots, and avoid active first-frame or shot-continuity guides. If resources are tight, testnative_guide_scaleat0.5. - Use the Advanced H3 conditioning node when you specifically need a setting picture, mask guide, or separate first-frame control; those controls are not exposed by this Simple node.
BFS Shot H3 Conditioning Common Errors and Solutions:
Native temporal alignment needs MODEL connected; route this node's MODEL output to the guider.
- Explanation: The experimental temporal mode cannot prepare its model without a
MODELconnection. - Solution: Connect a compatible model to
modeland connect this node’smodeloutput to your guider. Alternatively, select a non-temporalguide_mode.
Native temporal experiment requires 24 fps shots. Set the Planner output to 24 fps.
- Explanation: The shot frame rate does not meet the experimental mode’s exact 24 fps requirement.
- Solution: Set the Planner’s shot output to 24 fps and regenerate the shot data, or use a standard guide mode.
Native temporal mode requires first-frame and active shot continuity guides disabled.
- Explanation: Native temporal alignment detected an active first-frame or shot-continuity guide that it does not support. Although this Simple node does not expose a first-frame control, the incoming shot can still carry continuity settings.
- Solution: Disable the relevant continuity or head-guide settings in the Planner for this shot, then run conditioning again. Use a standard guide mode if continuity is essential.
Native temporal mode requires mask_guide off; use only the shot as native motion video.
- Explanation: The underlying conditioning path rejects mask-guide use in native temporal mode. The Simple node does not expose a
mask_guideinput and passes its off default, so this message is not expected from an unchanged Simple-node configuration. - Solution: Check whether the workflow is actually using the Advanced node or has been modified. Keep mask guidance off for native temporal alignment.
Colored Mark: a person is marked as <Picture n>, but the shot passes fewer pictures.
- Explanation: A Colored Mark configured in the shot points to a numbered reference picture that the shot does not provide.
- Solution: Add the missing picture to the shot or global reference list, or change the mark to an existing picture number.
Colored Mark: <Picture n> and <Picture n> have the same mark color.
- Explanation: Two different reference pictures have been assigned the same Colored Mark colour, making the intended person ambiguous.
- Solution: Give one picture a different mark colour in the Pictures & prompt settings.
Colored Mark needs a person selection (SAM 3 text/points or a mask video)
- Explanation: The incoming shot enables Colored Mark but has no usable person selection or crop information for placing that mark.
- Solution: Add a person selection through SAM 3 text or points, or supply a suitable mask video in the shot-planning workflow.
Audio is not included when with_audio is enabled
- Explanation: This is a troubleshooting symptom, not a quoted exception. The shown implementation attaches the shot soundtrack only when
with_audiois true andaudio_vaeis connected. - Solution: Connect
audio_vae, confirm the planned shot has audio, and keepwith_audioenabled.
Mask-only generation does not affect the intended subject
- Explanation: This is a troubleshooting symptom, not a quoted exception. Mask-only inpainting depends on a valid mask or person selection in the planned shot.
- Solution: Inspect the Planner’s mask and per-shot inpaint mode, or select forced mask-only generation after supplying a usable mask.
Normals do not produce the expected Colored Mark result
- Explanation: This is a troubleshooting symptom, not a quoted exception. The
normalsconnection is intended for shots whose Colored Mark style isnormal map. - Solution: Confirm that style is selected and that the connected normal images correspond to the shot’s frames; otherwise, leave
normalsunconnected.
