MiniMax H3 Image Studio v24.1.0

• Add chainable H3 LoRA nodes with individual filenames and manual strengths for generation and multi-reference editing.
• Include two optional LoRA slots in both starter workflows, bypassed by default. Duplicate nodes for longer chains; None/zero strength also loads nothing.
• Preserve upstream model bundles and cached branches; include all applied adapters and strengths in run information.
• Keep the older embedded Turbo shortcut compatible, and document how to avoid applying it twice.

MiniMax H3 Image Studio v24.0.0

• Replaced the starter interface with four nodes and two workflows using one shared model setup.
• Sample one H3 latent frame and include Fizgig's MIT-licensed temporal-group video-VAE decoder with attribution.
• Keep one optional Larry v4 adapter at 0.38 for the community still recipe, with manual steps.
• Remove frame counts, frame selection and hidden preservation prompts from the new interface. Keep megapixels, aspect ratios and advanced exact dimensions.
• Combine image-to-image and reference editing with up to nine dynamic picture inputs and no locked source frame.
• Cache prompt/reference preparation independently from seed and steps through ComfyUI's graph cache.
• Preserve old nodes as deprecated compatibility definitions and old examples as historical files. Exclude historical examples and the decoder converter from Registry packaging.
• Require ComfyUI 0.37.0 for the dynamic-input interface.

MiniMax H3 Image Studio v23.0.0

• Added paired four-step generation and reference-edit workflows with matching adapters; raised the eight-step entry points to about 1 MP.
• Added independent latent-slice decoding so image VAEs can decode stills without forcing one-frame sampling.
• Added a checked 500K image-decoder converter, including QKV head layout and SwiGLU gate-order conversion.
• Tested generation, edits, frame profiles, decoder alternatives, and the Qwen refiner at larger resolutions. Published measurements and unfiltered comparison sheets.
• Verified every explicit workflow input survives frontend import, save, and reopening; run tensor-based tests in CI instead of skipping them.
• Removed developer benchmark clients from the Registry package. The Registry flagged v22's local HTTP test script; runtime nodes do not need that script.
• Check Registry activation after upload; fail publication checks if the version is flagged or activation is unconfirmed.
• Removed a stale version label and documented practical defaults and experimental limitations.

MiniMax H3 Image Studio v22.0.0

• Added eight-step 768p generation and REF2VA editing workflows with matching Turbo adapters and separate sigma shifts.
• Added optional vision-encoder-only reference conditioning without VAE reference latents.
• Removed preservation wording that could contradict explicit edits at high source fidelity.
• Reduced the default reference resolution budget and clarified one-frame limitations.
• Kept existing workflow types and sampling recipes compatible; added regression checks for reference order, conditioning, and adapter schedules.

MiniMax H3 Image Studio v21.0.0

• Replaced the FLUX.2 Klein 4B refiner with a four-step Qwen Image Edit 2511 generative detail pass.
• Preserved the original H3 image dimensions instead of shrinking output to the model working canvas.
• Added a preservation-first Qwen edit prompt and full-resolution tone-lock path.

MiniMax H3 Image Studio v20.0.0

• Added a custom mode to Sampling Preset with sampler, scheduler, steps, denoise, H3 sigma shifts, and beta-scheduler controls.
• Kept documented and legacy presets immutable so existing workflows retain their exact recipes.

MiniMax H3 Image Studio v19.0.0

• Added an optional four-step FLUX.2 Klein 4B detail-refinement workflow for any H3 still image.
• Added Detail Tone Lock to restore H3 lighting and color while blending refined detail conservatively.
• Replaced broad restoration wording with a preservation-first prompt that protects eye state, identity, composition, and exposure.
• Added runtime-tested model, sampler, scheduler, prompt, and color-preservation guidance.

MiniMax H3 Image Studio v18.0.0

• Added dedicated one-frame text-to-image and image-to-image workflows using the experimental H3 image VAE.
• Changed one-frame I2I to use editable REF2VA source conditioning instead of an impossible FL2VA frame-0 anchor.
• Preserved the existing FL2VA keyframe path for every multi-frame I2I profile.
• Added runtime and release validation for all three one-frame workflows.

MiniMax H3 Image Studio v17.0.2

• Reduced the experimental single-image detail adapter from 1.0 to 0.5 after direct pose-transfer testing showed that full strength could suppress requested edits.

MiniMax H3 Image Studio v17.0.1

• Made explicit multi-reference trait assignments override Picture 1 preservation wording.
• Removed ambiguous still-image wording that could lock the result to Picture 1's composition.
• Lowered the reference-edit preservation default to 0.60 for visible pose and composition transfers.
• Corrected the official Turbo adapter download link.