APIXO's MiniMax H3 LoRA route generates video from prompts, frames and reference assets. Decide which workflow you need before collecting assets: an opening frame, an opening-and-ending pair or references that describe a new scene.
The current H3 LoRA Docs explicitly say the public interface does not accept custom LoRA configuration. This guide explains reference generation; it does not provide dataset-upload or custom-weight training steps.
Assign a role to each reference: Editorial workflow diagram, not a tested model output
A starting frame and a reference image have different jobs
In image-to-video, the first image establishes the opening, and a second image can establish the ending. Consider whether the two poses and compositions form a plausible transition.
In reference generation, images communicate a character, product, scene or style without necessarily becoming the opening frame. Treating these inputs as interchangeable can create the wrong expectation about the first frame.
Start at the MiniMax H3 LoRA model page and choose whether you are animating a source image or composing a new clip from assets.
Confirm the input workflow before submitting
Intended task: Animate a specific composition · Preparation: One starting image and a motion prompt · Submission check: The image defines the opening frame, not just a style
Intended task: Move between two states · Preparation: Compatible starting and ending images · Submission check: Check subject, framing and a plausible action path
Intended task: Combine a product, setting and camera reference · Preparation: References with explicit roles · Submission check: Check counts, mode and video/resolution compatibility
Intended task: Generate visuals from an audio reference alone · Preparation: Add qualifying visual input · Submission check: Audio alone does not satisfy this route's reference requirements
Intended task: Upload a training set or personal LoRA weights · Preparation: No such public step is exposed · Submission check: Do not treat the generation workspace as a training manager
The LoRA page discusses exploring at 480p and later generating at 768p. Those low-resolution attempts are drafts in a production sense; they do not establish a Seedance-style draft_task_id inheritance workflow. With video references, this LoRA route currently supports only 480p. You cannot simply resubmit the same video inputs at 768p; compare ordinary H3 when you need a higher-resolution video-reference workflow.
Keep the original image, crop and prompt. Improve occluded or unreadable source material first. If reference colors conflict, decide which image governs the result. Adding references or repeating a generation is not training and does not guarantee identity or product accuracy.
Give every asset a role
For a product shot, use a product image for appearance, a scene image for lighting and a short video for camera movement. Audio can contribute sound or rhythm, but cannot replace the required visual input.
An untested asset worksheet could include:
Product front view: appearance, color and packaging; exclude unrelated products.
Scene image: environment and light; exclude unwanted people or props.
Short video: movement and pacing; identify elements not to transfer.
Audio: sound direction; avoid unrelated dialogue and noise.
Describe the required contribution of each asset in the prompt. That is a planning method, not proof that the model isolates every requested attribute perfectly.
Check limits before submitting
The current route accepts up to 9 reference images, 3 videos and 3 audio clips, with 5–15-second output at 480p or 768p.
Video-reference input currently requires 480p on this LoRA route. Do not keep resubmitting 768p with video input when the current contract excludes it. Check the documented constraints before selecting another model.
An audio-only reference does not establish a valid reference-video request here. Provide the required visual input.
State what stays and what changes
Describe fixed appearance before action, camera and sound. Do not ask the model to reconstruct critical product details that are absent from the source.
Untested prompt example:
Use the product images for the bottle shape and color, and the scene image for background lighting. Use the reference video for a slow sideways camera move. Keep the product still and add no text, props or people.
Use the actual entry point's reference syntax. This example explains information roles without assuming one platform's special markers work everywhere.
Reduce conflicting inputs when diagnosing problems
For appearance errors, inspect source clarity, angle and subject count. For motion errors, check whether the video contains several actions or frequent cuts. Validate visual and sound requirements separately if they conflict.
Change one clear variable and retain all attempts. The LoRA name alone does not establish superior style control or character consistency over ordinary H3.
Check APIXO Pricing, including input charges rather than only the output-second starting price. The current model page charges reference assets as well as output.
API usage rates; Studio monthly credits are separate.
Review the generated clip
Check the subject, completed action, sound and usable ending. References communicate requirements; they do not guarantee exact product lettering or movement.
You do not need to train custom weights to use the currently exposed generation route. If you need higher output resolution or a different video-reference combination, compare MiniMax H3. Both routes support references; choose around the task's actual constraints.
Build with APIXO
Explore image, video, and audio models in one place.
