Blender Rendering Workflow

Read the shared execution guidance once per task for autonomous decisions, evidence cadence, and long-task continuity.

Treat a render as a reproducible deliverable, not a screenshot that happened to look acceptable once.

For architecture, daylight shafts or fabric-heavy scenes, read interiors and atmosphere. For photorealistic materials or environment lighting, use scanned materials and HDRIs: the bundled Poly Haven search/download tools acquire verified CC0 maps at up to 8K. Author the scene's important forms, but do not impose a from-scratch-only constraint that reduces realism.

Use the authored scene surface

Use blender_render_scene for final beauty renders and lighting diagnosis. Call it with inspectOnly: true first when camera names, dependencies or render settings are unknown. Then select existing cameras by exact name and optional frames; at most 12 camera/frame combinations are allowed. The default uses the active camera/frame, caps the longest edge at 1280 and Cycles samples at 64, and automatically selects an available GPU with an explicit recorded CPU fallback. Choose device: "cpu" or a specific backend when reproducibility requires it. Use a fresh output directory for each run. Inspect the inline PNG and render-manifest.json; errors are not successful evidence.

The standalone equivalent is bundled in this skill:

& $env:BLENDER_EXECUTABLE --background --factory-startup --disable-autoexec `
  --python-exit-code 1 --python "<skill-root>/scripts/render_scene.py" -- `
  --input scene.blend --output-dir renders/preview-01 --max-edge 1280 --samples 64

This preserves authored lights, world, volumes, cameras, materials and color management. blender_render_evidence deliberately replaces cameras and lighting; use it for standardized geometry checks, not as an interior's final image. Neither tool substitutes for authoring the scene in durable source.

Choose denoising for the task

Regular low-resolution renders use render denoising settings, not Blender's viewport preview settings. Select denoise: "preview" for look-development renders and denoise: "final" for final stills when residual noise is visible. The CLI equivalent is --denoise preview or --denoise final.

• Preview: prefer supported GPU OpenImageDenoise Fast with Fast prefilter; otherwise GPU OptiX when supported, then CPU OIDN Fast. Use a modest sample cap (for example 16-32) and 512-800px while checking composition and materials.
• Final: OpenImageDenoise High, Accurate prefilter, Color + Albedo + Normal, with GPU acceleration when supported and CPU fallback otherwise. Start from the scene's useful sample budget; do not raise samples just because it is final.
• denoise: "off" disables render denoising for an already-clean image or a raw noise/detail check. It does not disable an authored compositor denoiser.
• denoise: "preserve" is the backward-compatible default for deliberately authored settings and reproducible comparisons.

Inspect texture, fine edges, contact shadows and reflections at delivery size. When a surface looks smeared, compare a small raw crop or matching no-denoise preview before increasing samples. Use one denoising stage; an existing active compositor denoiser makes preview/final policy defer to the authored setup. Read the manifest's actual settings and fallback reasons. Denoising cannot repair inadequate geometry, missing texture or poor lighting; no pixel-based noise detector is implied by these policies. For animation, inspect consecutive frames for flicker before committing to a sequence.

The local mug/linen test favored GPU OIDN over CPU OIDN for inexpensive previews; this is a hardware/scene-specific result, not a universal speed ranking. See Blender sampling documentation for denoiser, prefilter and quality tradeoffs.

Establish the render contract

Before changing the scene, record:

• the intended audience, view distance, reference mood, and required story beat;
• still, turntable, image sequence, or video deliverable; resolution, frame range, frame rate, and output format;
• render engine, device assumptions, time/noise budget, and whether the result must match an engine or compositor downstream;
• required cameras, hero and diagnostic views, render passes, alpha, and color management requirements;
• which lights, world, materials, volumes, and compositor nodes are part of the authored result.

Keep these settings as named constants in the durable scene-generation script. Do not use a viewport screenshot as final evidence when the request calls for a rendered deliverable.

Author for repeatability

1. Name cameras, lights, world nodes, view layers, render settings, and output nodes semantically.
2. Set engine, resolution, frame range, sampling, denoise policy, transparent film, and color-management explicitly; never rely on a startup-file default.
3. Frame the subject with a deliberate focal length and camera height before increasing samples or adding post-processing.
4. Light for form: establish key, fill, rim/background separation, contact shadow, and exposure before cosmetic effects. Keep enough neutral evidence lighting to expose intersections and texture failures.
5. Use render passes and compositor nodes only when they improve the requested result. Keep the uncomposited beauty output available for diagnosis.
6. For turntables and animated output, render a short low-cost preview first; then lock camera and lighting before the final frame range.
7. For a render farm or external engine, emit a self-contained handoff manifest listing Blender version, engine, device assumptions, assets, fonts, cache paths, frame range, and output settings.

Iterate with evidence

For a normal asset handoff, include a polished hero presentation by default. Use a backdrop that separates the asset silhouette, softer key/fill lighting, controlled highlights, visible contact shadows, and framing that makes details readable. Match the material colors and avoid white-on-white or blown-out studio setups. Scale light power with squared scene dimensions when scaling distances and emitter sizes together.

Use the validation renderer's auto, neutral, dark, or light presentation as a starting point. Inspect the first image and adapt the preset or authored studio when contrast, translucency, emission, or textures require it. When two presentations fit distinct needs, deliver both with clear labels; do not make the user choose a preset before producing an initial useful result. Keep the fixed diagnostic views alongside the hero, and preserve explicitly requested art direction. A studio render does not replace required in-context shots.

1. Render a low-sample diagnostic frame for every required camera.
2. Open the rendered images, not merely file-existence logs.
3. Check composition, silhouette separation, exposure, specular control, shadow grounding, color balance, texture scale, noise/fireflies, clipping, and whether the requested material reads correctly.
4. Correct scene, material, camera, or lighting causes before tuning denoise, bloom, depth of field, or grading to hide them.
5. Render final stills or a bounded preview sequence. Open a contact sheet and, for motion, the encoded video or sampled frames.

Completion gate

Do not call a render complete until the source reproduces it in a clean Blender process and the final output has been visually opened. Report exact output paths, Blender and engine version, resolution, frame range, samples/denoise, color management, passes, render duration, and any machine-specific limits. Use $blender-agent-studio:blender-asset-validation for fixed multiview geometry evidence; that evidence supplements rather than replaces the art directed final render. For glass around an emissive subject, validate the final engine rather than the node graph alone. Keep transmission, alpha/blend behavior, refraction, exposure, and bloom controlled so the enclosure reads as glass and the emissive subject retains a distinct colored silhouette. An opaque or blown-white enclosure is a render failure even when its material is technically transmissive.