Character sheets with MiniMax-H3: prompt, references and result

.gitattributes

@@ -33,3 +33,9 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+examples/reference_1_front.png filter=lfs diff=lfs merge=lfs -text
+examples/reference_2_three_quarter.png filter=lfs diff=lfs merge=lfs -text
+examples/reference_3_back.png filter=lfs diff=lfs merge=lfs -text
+examples/reference_4_closeup_upper.png filter=lfs diff=lfs merge=lfs -text
+examples/reference_5_closeup_lower.png filter=lfs diff=lfs merge=lfs -text
+examples/result_character_sheet.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,137 @@
+---
+license: apache-2.0
+tags:
+  - minimax-h3
+  - hailuo
+  - character-sheet
+  - reference-sheet
+  - turnaround
+  - comfyui
+  - prompt
+language:
+  - en
+---
+
+# Character Sheets with MiniMax-H3
+
+**A prompt that turns a handful of reference photos into a clean 8-panel character sheet — angles on
+top, detail close-ups underneath — in a single 5-frame generation.**
+
+H3 is a video model, but it does not have to be used as one. Ask it for a *frozen* frame and it will
+happily lay out a reference sheet for you. Five frames is the node minimum, which is about 0.2
+seconds of video, so this costs a fraction of a normal clip. Export any frame and that frame is your
+sheet.
+
+The sheet is not a piece of concept art. It is a **documentation sheet of a design you already have**:
+the same object, from angles you do not have yet, on a neutral backdrop, at a consistent scale and
+under consistent light. That is exactly what you need to feed back into image-to-video work, where
+identity drift between shots is the usual problem.
+
+![result](examples/result_character_sheet.png)
+
+## What went in
+
+Five reference images of one armour set — three full views and two close-ups:
+
+| | |
+|---|---|
+| ![](examples/reference_1_front.png) | ![](examples/reference_2_three_quarter.png) |
+| front | three-quarter |
+| ![](examples/reference_3_back.png) | ![](examples/reference_4_closeup_upper.png) |
+| back | close-up, upper |
+| ![](examples/reference_5_closeup_lower.png) | |
+| close-up, lower | |
+
+Note there is **no side view among them**. The strict profile in panel 3 of the sheet is constructed
+by the model from the views that do exist.
+
+## What came out
+
+Eight panels in two rows.
+
+**Top row, four full views:** front · three-quarter · strict side profile · back. Same scale, same
+height, same camera distance, same light. Only the angle changes.
+
+**Bottom row, four close-ups:** head and shoulders · chest and waist · arm and hand · leg and foot.
+Each area appears once, shot straight on.
+
+## Settings
+
+```
+mode              Ref2VA
+length            5 frames        (node minimum; 5 % 17 == 5, valid without padding)
+resolution        1024 x 1024     (1.05 MP; H3's native canvas is a 768 short edge, so this is
+                                   slightly above it and still behaves)
+ref_image_size    max             (2048 short edge — keeps fine engraving readable; slower,
+                                   because reference tokens ride through every sampling step)
+sampler           res_multistep
+scheduler         simple
+steps             20
+guider            BasicGuider (H3 weights are CFG-distilled, there is no negative prompt field)
+```
+
+Because there is no negative prompt path, **every constraint is written as prose inside the positive
+text**. This matches MiniMax's own official ComfyUI templates, which end with lines like
+`No text, subtitles, logos or watermarks of any kind`.
+
+## The prompt
+
+Full text in [`prompt_character_sheet.txt`](prompt_character_sheet.txt). It follows the six-section
+Ref2VA format from MiniMax's official rewrite guide: `subject_definitions` · `summary` ·
+`retention_analysis` · `detailed_description` · `overall_soundscape` · `non_diegetic_music`.
+
+Three things in it do the heavy lifting.
+
+**1. Take it as it is.** A block stating that every plate, strap, buckle, spike, colour and finish
+keeps exactly what it has in the references, that nothing is redesigned, simplified, tidied,
+embellished or improved, and that this is documentation of an existing design rather than a new one.
+
+**2. A black mannequin instead of a wearer.**
+
+> The subject is worn on a PLAIN MATTE BLACK MANNEQUIN: a featureless display figure of uniform flat
+> black with no face, no eyes, no mouth, no hair, no skin texture, no visible seams and no joints,
+> smooth and completely non-reflective, present only to hold the armour in its proper shape.
+
+Without this the model invents a wearer, and that invented face is slightly different in every
+panel — which defeats the entire point of a reference sheet. With it, wherever the armour leaves the
+body open you get flat black instead of skin, and the panels finally agree with each other.
+
+**3. Every panel named, and named as distinct.** Early runs produced two identical head crops and two
+identical leg crops, because only two close-up references existed and the model simply repeated them.
+Naming the four areas explicitly and stating that each appears once and none repeats another fixed it.
+
+## Making it your own
+
+The prompt is written to be subject-agnostic. `<Subject 1>` is defined only as *the subject shown
+across the reference images* — no colours, no materials, nothing that could contradict what the model
+can already see in your photos. Swap the references and it works on a different armour set with no
+edits at all.
+
+For **mounts and vehicles**, delete the A-pose line.
+
+For **weapons**, keep the top row and rename the bottom four crops to the parts that carry the
+identity — grip, guard, blade, pommel.
+
+For **characters**, drop the mannequin block and let the person be the subject; keep everything else.
+
+For a **turnaround video** instead of a sheet — a slow full rotation on a neutral backdrop, from
+which you pull full-resolution frames — the same approach works with the panel layout removed. That
+gives you every angle at the full frame size rather than a quarter of it, at the cost of a real clip
+instead of five frames.
+
+## Notes from the runs
+
+- **Panel 3 first came out as a second three-quarter.** Writing *strict profile at exactly 90 degrees,
+  the far arm and far leg hidden behind the near ones, the face seen edge-on* fixed it.
+- **Close-ups came out doubled** until each was named and declared distinct.
+- **The invented wearer's face drifted between panels** until the mannequin replaced it.
+- A grid layout is a known weak spot for video models — panels bleed into each other. The bans on
+  repeated panels, overlapping panels, tilted panels and panel drop-shadows are there for that reason,
+  and the composition is declared a frozen frame where every frame is identical to the first.
+
+## Credit
+
+Reference armour renders are in-game captures from Diablo IV, used here purely as input photographs
+to demonstrate the technique.
+
+A ComfyUI workflow is work in progress and will be added here.

prompt_character_sheet.txt

@@ -0,0 +1,129 @@
+REFERENCE / DATA SHEET — assembled FROM the reference images
+Universal: armour · mount · weapon · character
+Mode: Ref2VA · ref_image_size: max · LENGTH = 5 FRAMES (the node minimum, 5 % 17 == 5, so it is
+valid without padding). Five frames at 24 fps is 0.2 s - effectively a still. Export any frame and
+that frame IS the sheet. Costs a fraction of a normal clip.
+
+The model is not inventing new angles here - it is LAYING OUT what the reference images already
+contain: the full views and the close-ups on the main parts.
+
+Canvas: 1920x1080 gives the top row 480x648 per panel and the bottom row 480x432 per crop.
+1024x1024 is tighter - use it only if you accept smaller crops.
+
+================================================================================
+
+subject_definitions:
+<Subject 1> is ONE single subject, shown across all the reference images from different angles and in
+different close-ups. They are all the same subject - not several different designs - and nothing is
+combined or mixed between them. Its every part, fitting, ornament, colour and material is exactly as
+it appears there. Only the subject itself is taken from those images: not their backgrounds, not
+their lighting, and not any other object or person visible in them.
+
+summary:
+[reference generation] The target video is a single still frame: a reference data sheet that
+lays out <Subject 1> as it appears across <Picture 1> to <Picture 5> - full views along the top and
+close-ups of the main parts along the bottom - against a neutral studio backdrop. The composition
+does not change for the length of the clip, so one frame can be exported and used as a reference
+sheet.
+
+retention_analysis:
+<Picture 1>: fully_preserved - the shapes, colours, materials and ornament of the subject are retained.
+<Picture 2>: fully_preserved - the shapes, colours, materials and ornament of the subject are retained.
+<Picture 3>: fully_preserved - the shapes, colours, materials and ornament of the subject are retained.
+<Picture 4>: fully_preserved - the shapes, colours, materials and ornament of the subject are retained.
+<Picture 5>: fully_preserved - the shapes, colours, materials and ornament of the subject are retained.
+
+detailed_description:
+The target video is a STILL IMAGE - a frozen reference data sheet in flat product-photography style, with no motion of any kind from the first frame to the last: a seamless
+neutral mid-grey background, soft even studio light from the front and both sides so every surface
+is readable and nothing falls into black shadow, neutral accurate colour, no colour cast, no
+stylised grading, no environment of any kind.
+
+WHAT IS TAKEN FROM THE REFERENCE IMAGES: the subject exactly as it already is. Every plate, panel,
+strap, buckle, spike, horn, edge, seam, layer and attachment keeps the shape, size and position it
+has there. Every colour, metal, cloth, leather, scale, gem and glow keeps its exact hue and finish.
+The proportions and the silhouette are the same. Nothing is redesigned, restyled, simplified,
+tidied, embellished, modernised or improved, nothing is added that is not on the references and
+nothing on them is left out. This is a documentation sheet of an existing design, not a new one.
+
+THE ONLY DIFFERENCE FROM THE REFERENCE IMAGES is who is wearing it: the subject is worn on a PLAIN
+MATTE BLACK MANNEQUIN: a featureless display figure of uniform flat
+black with no face, no eyes, no mouth, no hair, no skin texture, no visible seams and no joints,
+smooth and completely non-reflective, present only to hold the armour in its proper shape. No actor,
+no model and no person of any kind appears in the sheet. Wherever the armour leaves the body open,
+what shows underneath is the flat black mannequin surface, never skin.
+
+[Shot 1] The frame is a grid of panels on that grey background, separated by thin even neutral-grey
+gutters of equal width, with a matching margin around the outside. Every panel is a plain rectangle,
+flat to the camera, with no tilt, no perspective and no drop shadow.
+
+TOP ROW - FOUR FULL VIEWS, equal panels side by side, the whole subject in each, shown at exactly
+the same scale, at the same height, upright and level, centred in its own panel, lit identically,
+photographed from the same camera height and distance. Only the viewing angle differs:
+  1 - FRONT, facing the camera straight on
+  2 - THREE-QUARTER, turned 45 degrees to its own right
+  3 - SIDE, a STRICT profile at exactly 90 degrees: the camera sees only one side of the subject,
+      the far arm and far leg are hidden behind the near ones, and the face, if any, is seen edge-on
+      with the nose in silhouette. This panel must be clearly different from panel 2, not a second
+      three-quarter view.
+  4 - BACK, seen from directly behind
+
+BOTTOM ROW - FOUR CLOSE-UPS, equal panels side by side, each a tight crop on one main area of the
+same subject, filling its panel, on the same grey background and in the same light. The four areas
+are DIFFERENT from each other and each appears ONCE. No two of these panels show the same area, the
+same framing or the same crop, and none of them repeats another:
+  5 - HEAD AND SHOULDERS: the headpiece, the black mannequin head inside it, the neck and the
+      collar, front-on.
+  6 - CHEST AND WAIST: the torso plate, the belt line and what sits on it, front-on. No head in frame.
+  7 - ARM AND HAND: one whole arm from shoulder to fingertips, with the hand open and clearly visible.
+  8 - LEG AND FOOT: one whole leg from hip to the ground, with the foot clearly visible.
+Any angle or crop that the reference images do not already contain is built from the ones that do,
+keeping the same silhouette, plate shapes, layering and proportions.
+Each close-up is shot STRAIGHT ON, square to that part of the subject, filling its panel, large and
+readable, with no tilt and no oblique angle. The head close-up faces the camera front-on, not turned
+to the side. Each crop is cleanly framed on its area and is not cut through at an odd angle.
+
+It is one and the same subject in every panel, top row and bottom row alike, with the same colours,
+materials, proportions, ornament and number of parts. Nothing is added, removed, redesigned or
+restyled between panels; the back view belongs to the same object as the front view, and each
+close-up belongs to the same object as the full views.
+
+POSE - in the four full views the mannequin stands in a relaxed A-pose,
+feet shoulder width apart, arms held slightly away from the body with open hands and separated
+fingers, head level and facing along the body, the same pose in all four.
+
+This is a FROZEN FRAME. Nothing moves at all. The camera holds a static shot and does not move,
+zoom, tilt, shake or drift. The panels do not slide, resize, reorder, cross-fade or animate. The
+subject does not move, turn, gesture, breathe or change pose inside any panel. Any glowing or
+burning parts hold a constant even glow without flaring or flickering. Every frame is identical to
+the first: the shot reads as a single photograph, not as a video.
+
+No text, letters, numbers, captions, labels, titles, arrows, callouts, measurement lines, colour
+swatches, logos or watermarks of any kind anywhere in the frame or in the gutters, no second copy of
+the subject inside a panel, no repeated panel, no two panels showing the same area or the same crop,
+no extra panels, no empty panels, no overlapping panels, no other object
+or person, no actor, no human face, no eyes, no visible skin, no hair, no environment, no floor pattern, no cast shadow on the background, no camera movement,
+no cuts, no fades, no dissolves, no motion blur, no depth-of-field blur, no stylised or cartoon
+rendering, no dramatic or coloured lighting, no lens flare, no smoke or atmosphere.
+
+overall_soundscape:
+Silence.
+
+non_diegetic_music: N/A
+
+--- NOTES AFTER FIRST RUN (2026-08-10) ---
+Worked first try: 8 panels, grey background, no text, one subject throughout, top row and bottom row
+both correct in content. Two things to watch:
+  - panel 3 came out as a second three-quarter instead of a true profile -> the STRICT profile
+    wording above was added for that.
+  - the head close-up came out turned rather than front-on -> the straight-on wording was added.
+If the subject is an armour set with no chosen wearer, add to <Subject 1>: "worn on a neutral figure
+whose face is not the subject of this sheet" - otherwise the model invents a face and it will differ
+between panels.
+
+--- NOTES AFTER SECOND RUN ---
+Top row correct: front, three-quarter, TRUE side profile, back. The "constructed profile" wording
+fixed panel 3.
+Bottom row came out doubled - two identical head crops and two identical leg crops - because only
+two close-up references exist. Fixed by naming the four areas explicitly and saying each appears
+once and none repeats another.