game-dev-character-model

name: game-dev-character-modeldescription: Create a realistic, AAA-looking 3D character model for a game or web scene, from a description or a reference video/image, as a textured, rigged GLB standing in a relaxed pose in a lit three.js viewer. Covers designing an original character, the photoreal A-pose concept image, image-to-3D with Meshy or Tripo through Fil, rigging, posing in code, hangar-style lighting, headless capture, and the checks before showing it. Use when the user asks for a character model, a hero/heroine/NPC, "a model as good as X", a realistic human for a game, or to rig or pose a generated character.

game-dev-character-model

Makes a character that holds up at game distance (full body to half body) in about an hour for roughly 100 Fil credits.

Why it looks good

Not polygon count. The same character looked identical at 310k and 150k triangles. What carries the realism, in order:

1. The concept image. Image-to-3D bakes the picture into the textures, so a photoreal concept with real material detail (weave, stitching, scuffed edges, skin pores) gives a photoreal model. A weak concept cannot be rescued later.
2. 4K PBR textures with normal and roughness maps, so seams and scratches catch light without geometry.
3. Lighting. A dark scene with a cool key, a warm fill and two rim lights, ACES tone mapping and soft shadows. The same model under flat light looks like a toy.
4. A natural pose. An A-pose reads as an asset; arms down with a head turn reads as a person.

Steps

1. Design an original character

If the reference is a character from an existing game, film or show, do not recreate it and do not download ripped assets. Say so, then design an original one at the same fidelity: a different face, hair, silhouette and colour scheme. Take only the quality bar from the reference.

When the reference is a video on X, get it through api.fxtwitter.com/<user>/status/<id>, extract frames with the ffmpeg in imageio_ffmpeg, and look at full-resolution crops before deciding anything.

2. Concept image (about 27 credits)

Fil generate, kind image, model gpt-image-2.6-flare, 2:3, 2k, quality high. It quotes up to 105 credits for a batch of four but settled at 27. The prompt must contain all of:

• "Full-body character reference for a AAA game, photorealistic real-time render quality"
• Face, hair, build and outfit in concrete terms, with named materials and wear
• "standing straight facing the camera, relaxed A-pose with arms about 35 degrees away from the body, fingers slightly apart, feet shoulder-width, whole figure inside the frame with margin"
• "Front view, eye-level orthographic-style camera, no perspective distortion"
• "Even soft neutral studio lighting, no hard shadows, no rim light. Plain flat light grey background. No text, no logos, no weapons, no props."

Flat lighting matters: shadows and rim light in the concept get baked into the texture. Arms away from the body matter: arms against the torso fuse into it and cannot be rigged. Pick the output with the cleanest hands and the clearest gap between arms and body.

3. Image to 3D (60 credits)

Fil generate, kind 3d, wired from the concept node, rigging off.

• meshy-7.1 at resolution 4k is what worked: about 310k triangles, 50 MB GLB, 5 minutes.
• tripo-3.1 at ultra (77 credits) is finer on faces and hair, but on 12 Nov 2026 it failed with "403 not enough credit" on Fil's side. If it fails that way the credits come back; switch to Meshy rather than retrying.

A generate call can time out with an HTML "Inactivity Timeout" while the run still started. Check get_canvas before retrying, or it is paid for twice.

4. Rig (10 credits)

meshy-rig wired from the model node. It returns 24 Mixamo-style bones (Hips, Spine02, Spine01, Spine, neck, Head, Left/Right Shoulder, Arm, ForeArm, Hand, UpLeg, Leg, Foot, ToeBase), walk and run clips, and a mesh cut to about 150k triangles. There are no finger bones, so the hands keep their sculpted shape.

Keep the unrigged GLB too: it is the full-detail version.

5. Viewer and pose

Copy assets/viewer.html next to the GLBs (assets/heroine-rigged.glb by default, or pass ?m=). It has the light rig, floor, bloom, the four camera views, a turntable, and the pose code.

Pose by direction, not by Euler angles, because the bones' local axes are arbitrary. The viewer's aim(bone, child, worldDir) turns a bone so the segment to its child points along a world direction, and turn(bone, worldAxis, degrees) adds a twist. A relaxed stand is: upper arms [±0.30, -1, 0], forearms [±0.20, -1, 0.27], chest turned 5°, head turned about 10°. Leave the legs alone unless the feet are re-grounded afterwards. The breathing idle multiplies small offsets onto the posed quaternions.

Re-fit the arm directions for each character: hands must hang clear of hip pouches, holsters and coat hems.

6. Capture and check

The Browser pane has WebGL disabled, so capture headless:

node scripts/capture.mjs http://localhost:<port>/<folder>/index.html <out-dir>

Look at every image before sending. Check for:

• hands or forearms passing through the body or gear
• torn or collapsed mesh at the armpits and elbows
• feet on the floor, with a contact shadow
• a blown-out floor or face (lower exposure or the rim light, raise floor roughness)
• two views with the same file size, which means the render loop stalled and the frame is a duplicate

Send the A-pose and posed shots as a before/after pair, and say which view is weakest.

Known limits, say them plainly

• Face up close. Skin is flat and eyes look painted in a tight close-up. Fine from half body outward. A close-up hero needs a MetaHuman base instead, which takes many hours.
• One fused mesh. Hair, clothes and body are a single surface; nothing can be swapped or simulated without splitting it.
• No fingers. Gripping a weapon needs grafted rigged hands.

Traps

• page.waitForFunction polls on animation frames and never resolves on a heavy skinned mesh. Poll with page.evaluate in a loop.
• A second page.goto on the same page hangs while the model renders. Switch views by clicking the view buttons on one page.
• page.screenshot times out under load. Read the canvas with toDataURL (the renderer needs preserveDrawingBuffer: true).
• Set frustumCulled = false on the skinned mesh and recompute its bounding box after posing, or the grounding and culling use the bind pose.
• Check uptime first. With a load average above about 20 from other sessions, captures stall; wait rather than debug.