5tar5ystem MMH3 Character Sheet Builder 2.0A character sheet builder for MiniMax H3 in ComfyUI. You give it photos, videos and audio of aperson, say what each reference is for, and it renders a sheet of views you can feed back into aref2va workflow as one image.Required Custom node: <a target="_blank" rel="ugc" href="https://github.com/The5toryofthe5ecret5tar5ystem/5tar5ystem-MMH3-Character-Sheet-Builder">https://github.com/The5toryofthe5ecret5tar5ystem/5tar5ystem-MMH3-Character-Sheet-Builder</a><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/db2177e9-614e-4991-a150-1344f0028c11/original=true/db2177e9-614e-4991-a150-1344f0028c11.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/2cf7eb90-067b-4ecc-9c98-5724de2b32f1/original=true/2cf7eb90-067b-4ecc-9c98-5724de2b32f1.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/71c12e17-09e5-4ee6-98c4-42349a2d251b/original=true/71c12e17-09e5-4ee6-98c4-42349a2d251b.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/7d5125b9-f7ef-4b6d-8027-a68405c7d70c/original=true/7d5125b9-f7ef-4b6d-8027-a68405c7d70c.jpeg" />Version 2.0 changes two things: what a render is, and what the node looks like while you use it.Repo: <a target="_blank" rel="ugc" href="https://github.com/The5toryofthe5ecret5tar5ystem/5tar5ystem-MMH3-Character-Sheet-Builder">https://github.com/The5toryofthe5ecret5tar5ystem/5tar5ystem-MMH3-Character-Sheet-Builder</a>Release (v2.0.0): <a target="_blank" rel="ugc" href="https://github.com/The5toryofthe5ecret5tar5ystem/5tar5ystem-MMH3-Character-Sheet-Builder/releases/tag/v2.0.0">https://github.com/The5toryofthe5ecret5tar5ystem/5tar5ystem-MMH3-Character-Sheet-Builder/releases/tag/v2.0.0</a>Workflow, drag onto the canvas: <a target="_blank" rel="ugc" href="https://github.com/The5toryofthe5ecret5tar5ystem/5tar5ystem-MMH3-Character-Sheet-Builder/blob/main/example_workflows/5tar5ystem%20MMH3%20Character%20Sheet%20Builder.json">https://github.com/The5toryofthe5ecret5tar5ystem/5tar5ystem-MMH3-Character-Sheet-Builder/blob/main/example_workflows/5tar5ystem%20MMH3%20Character%20Sheet%20Builder.json</a>Workflow, direct download: <a target="_blank" rel="ugc" href="https://github.com/The5toryofthe5ecret5tar5ystem/5tar5ystem-MMH3-Character-Sheet-Builder/releases/download/v2.0.0/5tar5ystem.MMH3.Character.Sheet.Builder.json">https://github.com/The5toryofthe5ecret5tar5ystem/5tar5ystem-MMH3-Character-Sheet-Builder/releases/download/v2.0.0/5tar5ystem.MMH3.Character.Sheet.Builder.json</a>Workflow with the RefMod export appended: <a target="_blank" rel="ugc" href="https://github.com/The5toryofthe5ecret5tar5ystem/5tar5ystem-MMH3-Character-Sheet-Builder/releases/download/v2.0.0/5tar5ystem.MMH3.Character.Sheet.Builder.Plus.RefMod.json">https://github.com/The5toryofthe5ecret5tar5ystem/5tar5ystem-MMH3-Character-Sheet-Builder/releases/download/v2.0.0/5tar5ystem.MMH3.Character.Sheet.Builder.Plus.RefMod.json</a>The full manual is the README. The node's Help tab carries the same guide inside ComfyUI and checksyour install for the five model files it needs, with a tick or a cross against each one and thefolder it belongs in.// What 2.0 isTwo changes carry the version.1. One render makes the whole sheet, and it is the default. In 1.x every panel was its own shortclip, composited afterwards: five renders of 22 frames, five chances for the identity to driftbetween panels. In 2.0 one H3 clip draws the entire arrangement as a single image. The 5-frameminimum is enough because a sheet is a still. Frame 0 becomes the sheet, the other four frames arekept under <code>one_pass/</code>, and every panel is cut back out of the sheet at the box the prompt asked forand written where the per-cell pass writes its cells <code>cells/&lt;id&gt;.png</code>, <code>frames/&lt;id&gt;/f0000.png</code>). Thepicker, the Results tab and the RefMod export therefore work on a one-pass sheet without a secondrender. For a five-cell sheet that is 5 sampled frames instead of 110, and the panels agree with eachother by construction, since identity, lighting, palette and style cannot drift when one samplingcontext draws them all.2. The panel is a studio rather than a form. The layout is the one drawn in<code>web/mockups/panel-redesign-proposals.html</code> (proposal B): a narrow icon rail on the left holding theseven tabs, a one-line header with a status chip, two equal halves on the tabs that decide something(the stage and the control column), and a pinned footer whose primary action is Render.## What is new in 2.0* One-pass sheet as the default, with the per-cell pass unchanged and one switch away.* The panel: icon rail, one-line header, sheet and picture stages, control column, pinned footer.* Layout presets and quality presets as card rails. Each card draws what it does, so the layout cardfor "Hero + 4 panels" draws five boxes in the arrangement you will get.* Framings and expressions as glyph toggles instead of rows of checkboxes. Poses stay words.* A reference canvas: click a tile and it loads at the size of its column, with the paint surface,the Original / Blurred switch and the blur area chips on the picture itself.* Drag and drop per panel: drop a framing, a pose, an expression or a reference tile onto one paneland only that panel changes.* Clip blur, sound mute and hand-painted blur areas, on top of the picture blur 1.x had.* Per-cell clip export, with the audio H3 generated, at 24 fps.* The RefMod suite preset: four boards (hero sheet, expression board, SFW detail crops, NSFW detailcrops) built and queued as one job.* Your own presets, saved to ComfyUI's user folder and re-appliable to the next sheet.* Backdrops, including one borrowed from a reference's own location.* The live stream on its own tab, with the live strip that plays a looping clip while a cell isdenoised.* Settings drawn in the panel, grouped in three columns, with the node's own rows hidden so the nodestays short. Help carries the guide with the live file check.// The one-pass promptThe one-pass prompt is not the per-cell prompt repeated. It is written in the sections H3's owncaptions use: <code>subject_definitions</code> (which reference is the sole source of what), <code>summary</code>,<code>retention_analysis</code> (ownership and preservation, including that the identity picture's own outfit isreplaced when another reference owns the clothing), <code>detailed_description</code> (the layout),<code>overall_soundscape: None</code> and <code>non_diegetic_music: None</code>, because a sheet is silent.The layout is stated in words and in pixels. Each panel line carries its place both ways, forexample <code>Panel 2 of 5 (the middle column, upper half; x=1101, y=24, 1061x1034)</code>. On its own, pixelboxes were not enough: an early hero + 4 sheet came back as three full-height columns with two panelsmissing. The prompt also states that all panels are drawn and that none is merged, dropped, reorderedor stretched, that the full-body panels share one figure scale and one ground line, and that the sheetis a still, complete in the first frame and unchanged to the last, so the clip cannot drift into aturntable.Panel placement is still guidance, not a hard constraint. H3 treats the words and the boxes assemantic input, and it can reorder or merge panels. That is the honest description of any single-passmulti-panel prompt. When a panel has to be exactly right, or the sheet will be enlarged, turn theswitch off and render per cell.// What each mode gives up| | One pass (default) | Per cell ||---|---|---|| Sampling | one clip, 5 frames | one clip per cell, 5 to 425 frames each || Pixels per panel | its share of one canvas | a full render each || Frame picker, clips, audio, latent continuation | no | yes || Panel placement | guidance | each cell is its own render || Per-panel re-roll | no | yes, one cell with a new seed |## Resolution and memoryThe resolution control above the presets is the only thing that sets a size. One click sets two pairedsizes: the canvas a one-pass render draws, and the cell size a per-cell render uses. No preset toucheseither, so applying a layout cannot silently undo a 4K choice.| Button | Sheet short edge | Cell size | A 3:2 sheet renders | Latent tokens ||---|---|---|---|---|| 1080p | 1088 | 1024 | 1632 x 1088 | about 3.5k || 1440p | 1440 | 1024 | 2160 x 1440 | about 6k || 4K | 2176 | 2048 | 3264 x 2176 | about 14k |These are estimates from the pack's own numbers, not measurements: an int8 DiT is about 10.5 GBresident, the nvfp4 text encoder about 5 GB, and activations scale with the token count. Thecounter-intuitive part is worth knowing before you choose: a 4K one-pass render, about 14k tokens, ischeaper than one 2048px cell at 22 frames, about 18k tokens, because the one-pass grid is only fiveframes long.Rough guidance: 16 GB cards at 1080p with cell_size 768 and clip export off; 24 GB cards at 1440pcomfortably and 4K likely; 32 GB cards at 4K and cell_size 2048. A wrong guess costs a re-render,nothing worse.// Blur: pictures, clips and soundH3 conditions on every reference it is given at once and has no per-reference weight. A prompt canask for identity from the first picture and clothing from the second, but it cannot stop the modelreading a face that is present in the pixels. Blur removes it instead of arguing with it.* Pictures. Auto blurs a reference whose role does not claim the face, with per-tile on and off.Three areas: face, face and hair, whole head.* Hand-painted areas. For anything detection does not know about (a tattoo, a logo, a name tag, asecond person), open the preview and paint. Brush and lasso, stored as points on the referenceitself, so the painting is resolution-independent and travels with the workflow.* Reference clips. Sampled on 12 frames, tracked between them, blurred frame by frame and writtenback as a new file with its audio track copied. With the tile set to blur off, only what you paintedis blurred, which covers a logo or a watermark on a clip that keeps its framing.* Sound. There is nothing to blur in a voice, so a sound whose tile says mute on is written as asilent copy at the source's own rate, length and channels. The reference stays a usable sound in thegraph, it just carries no timbre. Auto keeps a voice, because a sound reference is usually added onpurpose.Blurred copies are cached in <code>input/h3_character_sheet/derived/</code> and keyed by the source's mtime andsize. Your original file is never modified.// Clips, continuation and the sheet on diskA per-cell render writes one clip per cell, h264 with the audio H3 generated alongside it, at 24 fps,into the sheet's own folder. H3 has no fps input: its joint video and audio latent is fixed at 24 fps,so the frame count is the duration and 22 frames is about 0.92 s. Switch it off while iterating and thethree nodes per cell and the encode disappear from the graph.Latent continuation anchors the last five frames of the previous cell at frame 0 of the next one,using ComfyUI core's H3 guide API. Auto chains only cells whose camera distance already matches, so aturnaround continues while a framing change stays crisp. It is a soft anchor rather than a splice:expect the same room, light and scale with a drifting pose. The frame picker skips the hand-overframes and ranks only the settled tail, because a chaining clip spends its first frames still in theprevious pose.Everything lands in one folder:```&lt;output&gt;/minimax_sheets/&lt;name&gt;/frames/&lt;cellId&gt;/f0000.png ... every frame the cell renderedcells/&lt;cellId&gt;.png the picked frame, or the panel sliced from a one-pass sheetclips/&lt;cellId&gt;_0000N_.mp4 the cell's clip, with its own audioone_pass/ the four frames of a one-pass sheet after frame 0&lt;name&gt;.png the composited sheet&lt;name&gt;.json manifest: spec, picks, sizes, warningspicks.json cell -&gt; {mode, index}report.txt human readable last run```// Presets and the RefMod suitePresets come in two axes. The layout axis says what to draw: Hero + 4 panels, Expressions 2x3,Turnaround, the two Closeups 2x2 boards, or the RefMod suite. The quality axis says how to sample it:one-pass sheet, per-cell renders, or max identity fidelity. A preset writes the node's own widgets andbuilds its cells, so picking one and pressing Queue is the whole job. Your own presets save the samethings to <code>user/default/h3_character_sheet/presets.json</code>, and each one computes its own list of thesettings it moves off a fresh node.The RefMod suite is a layout preset that renders four boards in one queue: the hero sheet, theexpression board, the SFW detail crops and the NSFW ones. Each board is a complete sheet folder underthe run name, and a <code>suite.json</code> records which boards belong together.## Exporting the sheet as a RefModA character sheet is a multi-view identity board, which is what a RefMod wants: a VAE latent that H3attends to like a real reference, at a fraction of the tokens. The <code>H3 Sheet -&gt; RefMod</code> node (from[ComfyUI-MiniMaxH3Mod](<a target="_blank" rel="ugc" href="https://github.com/Luisacaotica/ComfyUI-MiniMaxH3Mod">https://github.com/Luisacaotica/ComfyUI-MiniMaxH3Mod</a>)) writes one bundle withthe picked cells stacked, the composite as a second member, and a voice member taken from the sheet'sown reference audio, the soundtracks of your reference videos, or a cell's generated audio. FullReference stores the real encode so a face survives; Compressed Reference pools it into a tiny gridthat carries concept rather than identity. The node's report prints each member's token count, becausethe tokens are what the export costs at sampling time, not the file size.// Install1. ComfyUI-Manager, Custom Nodes Manager, search MMH3 Character Sheet (publisher <code>5tar5ystem</code>), thenInstall and restart ComfyUI.By hand:<code>cd ComfyUI/custom_nodes &amp;&amp; git clone https://github.com/The5toryofthe5ecret5tar5ystem/5tar5ystem-MMH3-Character-Sheet-Builder.git ComfyUI-H3-Character-Sheet</code>2. Restart ComfyUI, then hard refresh the browser with Ctrl+Shift+R.3. Load the ready-made workflow. It wires the H3 ref2va checkpoint, the Qwen3-VL text encoder, bothVAEs and a SaveImage, and pre-builds a five-cell matrix.No other custom node pack is required. Renders use ComfyUI core H3 nodes; this pack adds the sheet.## Models the pack needs| What | File | Where ||---|---|---|| Diffusion model | a TURBO beta5 H3 checkpoint, for example <code>10Eros_Max_h3_TURBO-hybrid_beta5_w4a8_14gb_optimized.safetensors</code>, or the official repack <code>minimax_h3_ref2va_pruned_int8_convrot.safetensors</code> | <code>models/diffusion_models/</code> || Text encoder | <code>qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors</code> | <code>models/text_encoders/</code> || Video VAE | <code>minimax_h3_video_vae_fp16.safetensors</code> | <code>models/vae/</code> || Audio VAE | <code>minimax_h3_audio_vae_fp32.safetensors</code> | <code>models/vae/</code> || Face detector, optional and only for the blur | <code>face_yolov8m.pt</code>, plus the <code>ultralytics</code> and <code>opencv</code> python packages | <code>models/ultralytics/bbox/</code> || Tiny VAE, optional and only for the live strip | a <code>taeh3</code> TAEHV decoder | <code>models/vae_approx/</code> |Downloads: Comfy-Org's H3 repack at <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main">https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main</a>, communityTURBO checkpoints at <a target="_blank" rel="ugc" href="https://huggingface.co/TenStrip/10Eros-Max/tree/main">https://huggingface.co/TenStrip/10Eros-Max/tree/main</a>, the face detector at<a target="_blank" rel="ugc" href="https://huggingface.co/Bingsu/adetailer/tree/main">https://huggingface.co/Bingsu/adetailer/tree/main</a>.Steps: 8 with a TURBO checkpoint, which is the node's default, or 20 to 30 with a non-turbo one. Witha TURBO file, do not also load a turbo LoRA, and skip cache or Spectrum nodes on reference runs.// A first sheet1. Drop your references into the References tab and type a role under each one. Fill in the Charactertext with who this is and Suppress with what must not appear, for example `text, watermark, logo,extra people`.2. Pick a layout card. The full-sheet presets build their five cells for you, or tick views, poses andexpressions yourself and press Build cells.3. Read the Prompt tab. That is the text the render will use, one section per panel in one-pass mode.4. Press Render in the panel footer. The Preview tab shows the frames as they are denoised.5. In Results, click a thumbnail to choose a different frame for a cell and press Re-compose. That onlyre-composites, so it costs no GPU time. In one-pass mode the panels are already the frames.// Two things worth knowingUpgrading from 1.x. The <code>single_pass</code> widget is on by default. A workflow saved before 2.0 has novalue for it, so it opens in one-pass mode. The One pass and Per cell segment in the Cells tab, or thesingle_pass switch in Settings, puts it back to one render per cell. Nothing else in the payloadchanged shape, and old sheet folders still list and re-compose.One good identity reference beats six mediocre ones. Extra pictures are for clothing, angles andobjects. Several different faces is the most common way a sheet goes wrong, which is also what theblur is for.// Credits and licenseGPL-3.0. Sheets are rendered by ComfyUI's own MiniMax H3 nodes; this pack adds the panel, the cellmatrix, the reference roles, the blur, the frame picker and the compositor. The sheet modules wereoriginally written inside ComfyUI-MiniMax-H3-Motion-Director and are carried here under the samelicense. Full credits are in <code>NOTICE</code> and the README. Feedback and issues are welcome on GitHub.

5tar5ystem MMH3 Character Sheet + RefMod Builder

v1.2.3: Your voice reference now fully conditions the voiceThe RefMod export voice member used to be the ~1s clip H3 generated with a cell, while the audio you had actually put in the Builder's References tab, the file that conditioned the whole sheet render was never used. That is the wrong way round, and it is measurable: 0.95s of H3's audio codec is 76 rows out of a ~15,000-row packed sequence the model attends over, ~0.5%. A/B measured on a real bundle: <code>voice 1.0</code> against <code>voice 0.0</code> came out at noise level, because the reference was working, it was just a rounding error.Two changes, no new widgets, nothing to re-wire:<ol><li><code>voice_cell = -1</code> (the default) is now a ladder: the sheet's own reference audio (all of it, joined) → else every exported cell clip joined → else a connected <code>AUDIO</code>.</li><li>The report prints every member's share of the rows, and tells you the <code>copies</code> count that would put a thin voice on the map.</li></ol>Same sheet, same wiring, one queue:<pre><code>voice: the sheet's reference audio (3 files joined) as 'hero_voice'
 - hero_views (video): 2000 tokens, 54.8% of the bundle, source stack
 - hero_voice (audio): 30.00s / 2400 tokens, 45.2% of the bundle, source audio
</code></pre>Added<ul><li>The export report gives every member its share of the bundle's rows. Rows are what a reference is worth — everything in a bundle is packed into one sequence the model attends over, and the video being generated adds its own rows on top of the bundle's. A member that reads 1% is a rounding error next to a target's thousands, and the report now says so.</li><li>A <code>copies</code> hint when the voice is thin. Under 2% of the bundle the report names the number their loader's <code>copies</code> widget needs to lift the voice into range (<code>copies 3</code> on a 0.5% member = 1.6%), or says so when even ten copies would stay thin. The 2% threshold is applied to the printed number, so a member reading <code>2.0%</code> is never called thin on the next line.</li><li><code>voice_cell = -1</code> walks a ladder: reference audio → cell clips joined → a wired <code>AUDIO</code>. <code>voice_cell = 0</code> still means "no sheet audio, use the connected clip", and <code>n</code> still forces the nth cell's clip (<code>&lt;name&gt;_voice_cellN</code>).</li></ul>Changed<ul><li>The voice member defaults to the sheet's own reference audio — the files in the Builder's References tab, recorded in the run's manifest (<code>spec.refs.audios</code>). Disabled rows are skipped, so ticking off one of three clips in the panel is how you choose. Measured live: the same sheet that produced a 0.95s / 76-token voice now produces 30.00s / 2400 tokens (31x), and up to 64s fits the 5120-token encoder budget.</li><li>Every exported cell clip is joined instead of just the first one — one member, <code>&lt;name&gt;_voice_cells</code>, resampled to H3's 32 kHz and concatenated as samples before the single encode. H3's audio latent clock is a fixed 40 frames per second, so joining the latents of separate encodes would cut the timeline mid-frame; joining the waveforms cannot. Verified: two clips → 1.88s / 150 tokens, in order.</li><li><code>voice_seconds</code> is documented as what it is: a ceiling, not a target (truncated from the front, never refused). A short reference is the usual reason a bundle feels inert.</li></ul>Fixed<ul><li>A reference file that has moved is reported and skipped, not fatal. A manifest row whose file is no longer on disk names the file in the report (<code>... 'voice.wav' is not on disk (ComfyUI's input folder) - the ladder cannot use it</code>) and falls through to the next rung instead of failing an export whose appearance members are perfectly fine.</li><li>The report names whatever the ladder passed over (<code>voice: not used - its 2 exported cell clip(s), the connected audio (...)</code>) so the override is discoverable rather than mysterious. Updated to export Character Sheets and RefMod using the generated characters + Audio</li></ul>

Files with SHA256: ed455d774bd9eb5ab756591b3e521574521d131cdbb24e00e251a254e20cc8f7