ZIP FILE INCLUDES GUIDES, NODES, how to set it up# Using the Studio## Choose the branchT2V: describe the initial composition, actions, camera and sound. No image anchor is required.I2V: load your starting image into the image input. Describe how that exact composition begins moving. Avoid requesting an immediate replacement of the image. The opening clip receives the first-frame alignment instruction; later clips should continue the incoming action.REF2V: load the intended reference images and keep their connected aliases consistent with the prompt. State what each reference contributes: subject identity, clothing, environment or composition. A reference image is not automatically the first frame. Do not reference assets that are not connected.## Set up a runUse the Production Controls for the model, matching base LoRA and sampling preset. Keep the release defaults for your first test. Choose a fresh run name when changing the story or inputs so saved scenes from an earlier run cannot be mistaken for current results.Review the plan in the Studio plan/editor nodes before queueing. Each plan scene produces a clip; it does not automatically imply a camera cut. Keep scene IDs stable within a project and check durations, seeds and scene order. For continuity, describe the end state of one scene as the starting state of the next. Use the review nodes to inspect the plan before committing to a long render.## Write or rewrite promptsManual mode uses your authored plan. Auto mode rewrites the plan through the optional local writer. Review its returned draft before applying it. If a rewrite fails, correct the error and generate a fresh draft; do not apply an unrelated old draft.T2V and I2V use integrated_multimodal_description, overall_soundscape and non_diegetic_music. Their visual description begins with [Shot 1]. I2V also has an application-owned opening image alignment instruction.REF2V uses subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape and non_diegetic_music. Its detailed description has style prose before [Shot 1]. Keep reference definitions and retained roles consistent.Every dialogue span needs its own actual language marker, including after a cut:(S1) says quietly, &lt;d&gt;[English] The cup is still warm.&lt;/d&gt;(S1) continues, &lt;d&gt;[English] Leave it here.&lt;/d&gt;Use [Slovak] for Slovak speech, or What ever Language and preserve the original words. Do not literally write [Language]. Keep speaker identity and delivery outside the dialogue tags. Do not add music unless wanted; use N/A in the music field when absent.## LoRAs and attentionBoth PixaromaLoraLoader and H3PowerLoraStack are included before the attention selector. Start with optional style LoRAs disabled or at zero, then add one at a time. The matching turbo LoRA is separate from optional creative LoRAs; keep it paired with its intended model and step count. Avoid loading the same style LoRA twice.The boolean attention switch selects Veda when ON and Sparse Attention when OFF. These are alternative paths. Confirm the Veda predictor matches the branch. The public sparse option uses sol-attn, not VSA/FastH3 weights.## What happens after generationSource scenes are saved and a base master is assembled. The finisher reads each delivered scene separately, in batches of at most max_frames_per_batch. Each batch passes through RelayVSR 4x, DLSS 5 at 1x and RTX HDR PQ, then is encoded as 10-bit HEVC. Frame tensors are released between batches.After processing, the batches are joined by stream copy. Original audio is copied once from the base master. Finished master path shows the result. Per-batch files and progress.json remain under output/H3_Pixel_Finish/&lt;run&gt;. Retained batches help diagnosis; automatic resume is not promised.## Finish an existing videoOpen 04_Finish_Existing_Video.json. Enter a full source video path or a path relative to ComfyUI/input. Set a fresh output name. The same processor splits this single input into bounded windows, then joins the finished result. Test a short SDR clip before a long video. This workflow performs an HDR conversion; it is not a verified HDR-to-HDR preservation path.

MInimax H3 Studio, Pixel Upscale Optimized for NVIDIA ONLY

Workflow Update – New Features, Better Prompts &amp; More Control!I've made several improvements to the workflow, added more customization options, and expanded support for different checkpoints. What's New?<ul><li>Longer Auto-Generated Prompts – Improved the built-in prompt generator to produce longer, more detailed prompts.</li><li>Improved Draft Writer – Corrected the output formatting for more consistent prompts.</li><li>Expanded Ref2V Support – Added more reference options for reference-to-video generation.</li><li>Turbo LoRA Loaders – Added dedicated Turbo LoRA loaders for non-Turbo checkpoints.</li><li>Regular LoRA Support – Integrated the regular LoRA loader from the Pixorama pack for standard LoRAs.</li><li>Attention Optimizations:<ul><li>VEDA Sparse Attention for non-Turbo checkpoints.</li><li>Sparse Attention for Turbo checkpoints.</li></ul></li><li>Optional HDR – HDR processing can now be bypassed.</li></ul>⚠️ Important: Turbo LoRAs &amp; Color DriftBe aware that Turbo LoRAs can introduce color drift when used with base checkpoints.You can adjust the AV Color Drift settings to reduce the effect, but some color shifting may still be visible.I recommend using Turbo LoRAs only if you're comfortable with this trade-off.🎬 My Usual Generation SettingsFor anyone curious about my personal settings:<ul><li>Resolution: 1.0–1.2 MP</li><li>First scene: 10 seconds</li><li>Following scenes: 7 seconds each</li><li>HDR: Optional, depending on the result I'm aiming for.</li></ul>These are simply the settings I usually work with, so feel free to experiment and find what works best for you. Questions or Feedback?If you have any questions, run into issues, or have suggestions for improvements, feel free to leave a comment below!Happy generating

Files with SHA256: aa3215231e5f658d98c3efab52372aa80a2560740444d613bf6aac5fccb7d93a