MiniMax H3 Character Swap V1.5 is Akatz Labs' updated adapter for replacing a person in a source performance with a character supplied through an image. The useful question is which checkpoint and workflow to use, especially if you already have a V1 installation. V1.5 is an adapter release, so downloading its small weight file does not supply the H3 base, text encoder, VAEs, or complete application.
This guide follows the author's current documentation, reviewed October 17, 2026. It is not a new H3Vid generation test. The cover is an AI-generated concept illustration, not a V1.5 result.
Choose the release deliberately
The author's model card recommends h3_character_swap_v1.5.safetensors for new swaps, starting at strength 1.0. V1 remains available for older projects and comparisons. The selected examples show stronger identity retention in some shots, while the card still describes the model as experimental. Treat the side-by-side video as selected evidence, not a universal success rate.
The V1.5 provenance record resolves a naming trap: embedded metadata uses an internal V2 name and an original 2,250-update budget, but the final release records 1,000 completed additional optimizer updates. It was initialized from V1. Use V1.5 in place of V1 rather than stacking both character adapters. The separate Turbo adapter serves a different purpose and belongs to the documented tool recipe.
Keep the filename, revision and checksum with your project. A screenshot saying “V1.5 selected” is weaker evidence than knowing which bytes were loaded, particularly when several renamed copies exist in a shared model directory.
Pick the app or standalone graph
The local workbench provides shot preparation, mask correction, segment prompts, takes and dependent continuation. Its current documentation makes V1.5 the default for new shots while preserving saved selections in existing shots. Opening an old project therefore does not prove it has upgraded.
Start from the repository's source installation guide. The public README says Docker alpha images are still being prepared and provides no published image digest. Do not turn a local development tag into an installation recommendation. Its qualified source target is a Linux NVIDIA machine; WSL2 is a separate qualification target.
For an existing ComfyUI installation, read the standalone workflow guide. That downloadable workflow is explicitly documented around V1. Loading it is not sufficient to select V1.5: inspect the character adapter node and replace the V1 file deliberately, while retaining the separate Turbo recipe. Keep a copy of the original graph for comparison.
Prepare one tractable shot
Begin with a short continuous performance and a reference that clearly shows the intended face, hair and clothing. Split camera cuts beforehand. If two people cross, identify the subject through masking before spending time tuning identity prose.
The standalone guide's practical sequence is to select the detected person, inspect the expanded green-mask preview at several positions, then review the performance prompt. A successful execution can still select the wrong person or retain too much original appearance. Check a front view, profile, an occlusion and a full-body frame rather than judging only the opening image.
An illustrative prompt to adapt is:
Replace the selected performer in Video 1 with the explorer in Picture 1. Keep the explorer's short dark hair, green coat and brass shoulder clasp. Follow the performer's turn, hand gesture and gaze. Preserve the other person and the room.
Translate those labels to the exact numbered reference syntax expected by your graph. An optional scene anchor has its own reference position; it should not accidentally become the character identity image.
Keep the recipe and audio decisions separate
The author documents a Ref2VA INT8 base with the character adapter and FL2V Turbo8, eight steps, CFG 1, er_sde and normal. This is a specific execution recipe. Changing a loader, base mode or sampler changes the experiment, even when the public name still says H3.
Source-audio conditioning and soundtrack preservation are different operations. The tool exports the original soundtrack; this does not establish perfect generated speech or exact lip synchronization. When troubleshooting, ask separately whether the right waveform reached conditioning, whether the visible mouth followed it, and whether the final export contains the intended source audio.
Review continuation before a long render
The app lets later segments depend on preceding takes. Choosing a different earlier take can mark dependent continuations as needing refresh. Save prompts and compare current takes before assembling the shot; otherwise an apparently random identity jump may be an outdated dependency.
Inspect face shape, costume details, hands, overlapping people and boundary motion through full playback. The author's validation notes also describe experimental quality and memory limitations. A 4090 example is not a guarantee that every one-megapixel recipe fits without offloading.
H3Vid's online character-swap page offers a separate reference workflow. It should not be understood as this local V1.5 workbench or its exact masked recipe. For V1.5, start with the author's pinned resources, prove one manageable shot, and only then extend the project. Models and application components retain their respective licenses.