ComfyUI Node: MiniMax H3 Upscale Reference + Prompt Override

MiniMax H3 Upscale Reference + Prompt Override:

The MiniMaxH3UpscaleReferencePromptOverride node is designed to enhance the upscaling process by allowing you to override the default prompt and reference settings during the second pass of video processing. This node is particularly useful for refining the appearance and details of upscaled video content without altering the original motion or camera instructions. By providing a mechanism to input a custom text prompt and tagged references, it enables more precise control over the visual output, ensuring that the upscaled video aligns with specific artistic or technical requirements. This node is essential for artists looking to maintain consistency in visual style while leveraging advanced upscaling techniques.

MiniMax H3 Upscale Reference + Prompt Override Input Parameters:

target_video_latent

This parameter represents the actual 24-channel pass-2 video latent. It is crucial for ensuring that the match-sized picture references and their Qwen presentation are accurately rebuilt for the H3 canvas. This latent is essential for maintaining the integrity of the video content during the upscaling process.

video_vae

The video_vae is the original MiniMax H3 video VAE used to re-encode match-sized picture references at the pass-2 canvas. It is important to use the correct VAE to ensure that the video content is accurately processed and that the upscaling results are consistent with the original video quality.

prompt_override

This optional parameter allows you to input a custom text prompt for the second pass. If left blank, the node will reuse the compiled generation prompt. It is recommended to use a concise appearance or detail prompt to avoid repeating source motion or camera instructions, which can lead to inconsistencies in the final output.

tagged_references

This parameter allows you to input an optional pass-2 Tagged Ref line from the Upscale Reference + Prompt Override. When connected, it replaces the cached reference payload and is compiled with the prompt_override. This feature is useful for artists who need to ensure that specific reference details are maintained during the upscaling process.

state

The state parameter represents the current upscale state. It uses the selected source branch, scene, prompt, and generation fingerprint to discover references. This parameter is essential for maintaining the continuity and consistency of the upscaling process.

clip

The clip parameter is the MiniMax H3 text encoder used to rebuild the cached Ref2VA presentation at pass 2. It plays a critical role in ensuring that the text-based references are accurately processed and integrated into the upscaled video content.

missing_cache

This parameter determines the action to take when caches are missing. The options are text_only, which falls back to text if recovery is unavailable, and error, which reports the missing source or VAE without dropping references. This setting is important for handling incomplete or missing data during the upscaling process.

motion_ref_mode

The motion_ref_mode parameter defines the pass-2 motion-reference policy. The default setting removes both native video-ref latents and their Qwen presentation while keeping paired audio. This parameter is crucial for managing how motion references are handled during the upscaling process.

audio_vae

The audio_vae is needed only when the connected override registry activates standalone or paired reference audio. It ensures that the audio component of the video is processed correctly during the upscaling process.

override_ref_image_size

This parameter controls the picture sizing for cached, rebuilt, or connected references. The options are inherit, match, and max, with inherit preserving the saved match/max policy. This setting is important for ensuring that the visual content is scaled appropriately during the upscaling process.

override_reference_policy

This parameter defines the prompt/tag validation for connected pass-2 Tagged references. The default setting is strict, which ensures that the references are validated according to the specified policy. This parameter is essential for maintaining the integrity of the reference data during the upscaling process.

MiniMax H3 Upscale Reference + Prompt Override Output Parameters:

compiled

The compiled output represents the final text prompt used during the upscaling process. It is either the custom prompt provided or the compiled generation prompt from the cached data. This output is crucial for understanding the text-based instructions that guided the upscaling process.

MiniMax H3 Upscale Reference + Prompt Override Usage Tips:

• Use the prompt_override parameter to refine the visual details of your upscaled video without altering the original motion or camera instructions. This can help maintain consistency in the visual style.
• Ensure that the video_vae parameter is correctly set to the original MiniMax H3 video VAE to avoid inconsistencies in the upscaling results.

MiniMax H3 Upscale Reference + Prompt Override Common Errors and Solutions:

ValueError: Pixel Current Scene requires Upscale Adapter backend=pixel.

• Explanation: This error occurs when the backend configuration is not set to pixel for the Pixel Current Scene.
• Solution: Ensure that the backend configuration in the state is set to pixel to avoid this error.

ValueError: Pixel Current Scene decoded %s; expected %d RAW RGB frames at %dx%d.

• Explanation: This error indicates a mismatch between the expected and actual dimensions of the decoded video frames.
• Solution: Verify that the correct H3 video VAE is being used and that the video content has not been trimmed or altered in a way that affects its dimensions.