Vulkan Offline for Daz Studio 6

Experimental Windows x64 offline renderer for Daz Studio 6 using Vulkan compute, hardware ray queries and a progressive path tracer. The goal is a drop-in replacement for Iray on GPUs Iray cannot use; the target is "similar enough" to Iray, not identical (SCOPE.md). The current implementation is not yet production-ready.

Target and current state

• Windows 11, AMD Radeon RX 7900 XTX, 24 GB VRAM (vendor-neutral Vulkan; other GPUs untested).
• Daz Studio 6.25.2026.14722; compile against SDK 6.25.2026.13900 in Daz Studio 6.25+ BETA SDK.
• Daz Studio 6.25+ BETA SDK Latest is reference-only: its SDK 6.25.2026.19400 requires a newer host. Its existing Qt development installation can be used without compiling against its SDK.
• Qt 6.10.3 MSVC 2022 x64, including Core5Compat. Vendor SDKs are unchanged.
• Current release: 0.1.1.91 (November 4, 2026). Against Iray (whole image): two-character scene 1.00, emissive-lit living room 1.00, render_test4 (strand hair, PBRSkin makeup, ghost lights) face 1.05 / body 1.00, render_test5 (night street: instances, area lights, Matte Fog, twilight Sun-Sky; six cameras) 0.97-1.00, render_test6 (geografts, sunlit room; two cameras) 0.99, render_test7 (pool room with large windows; four cameras) 0.96-1.01; visible noise at or below Iray's in the living room, 1.0–1.5× in the character scene at equal samples. Sampling at 512² 512 spp: characters 19.2 s, render_test4 face with dense strand hair ~59 s (host check 0.1.1.82, bounce limit 32). Renders are bit-identical run to run (render_test6's varying Daz hair aside). Status, out-of-bar regions and next steps: HANDOFF.md. Host tests run through the harness in ../../daz tools (headless Daz, Iray reference cache).

Milestones (details in VALIDATION.md and its archive):

Release	Change
0.1.0.19–20	Refractive light connections (manifold NEE) through solid glass
0.1.0.21	Learned path guiding (off by default; never a host gain; removed in 0.1.0.87)
0.1.0.23–28	Iray-style tone mapping, PFM output, Iray brightness/dome calibration, Draw Dome
0.1.0.29	Iray Uber glossy layering; final continuation ray traced (path length N ≈ Iray N+1)
0.1.0.30–31	Owen-scrambled Sobol sampling (always on)
0.1.0.32–35	Paged and grayscale-packed textures; PBRSkin with subsurface scattering
0.1.0.36	Glossy anisotropy; PBR-mode refraction reflection weight (eyes)
0.1.0.37–38	Adaptive GPU batching, cached readback, parallel texture decoding
0.1.0.39–40	Non-thin-walled Iray Uber translucency (hair cards)
0.1.0.41–42	Texture mipmaps with camera-driven resolution selection
0.1.0.43–46	Iray simple_glossy lobe, Iray curve-layer layering, Top Coat, Dual Lobe Specular, weighted lobe selection
0.1.0.47	Faster subsurface random walks (3.3× on a skin benchmark)
0.1.0.48–51	Iray Uber PBR Specular/Glossiness (layered base, coats, Glossy Specular as F0)
0.1.0.52	Emission blackbody temperature; map-less Iray dome renders black
0.1.0.53–54	Open Image Denoise (HIP GPU); DAZ Script access to render settings
0.1.0.55–56	Per-object geometry log; in-session texture cache (repeat renders ~2× faster)
0.1.0.57	Iray-calibrated photometric efficacy (emitter probe 1.000)
0.1.0.58–60	Review fixes; adaptive (learned, visibility-aware) light selection; emitter-guided paths default off
0.1.0.61–63	Iray MDL bump (file_bump_texture); Iray Uber Base Color Effect and SSS Mode bindings
0.1.0.64–65	Per-channel display firefly filter (red skin speckles); albedo-weighted subsurface channels
0.1.0.66	Daz gamma 2.2 colour decoding; mapped Iray Uber metal tint; Daz light units (point-light probes)
0.1.0.67	Iray flux-unit emission (lm, W): luminance = flux / area per side
0.1.0.68	Iray Visible to Primary Rays and Ghost Light Factor per node
0.1.0.69	Strand hair as tapered tubes (provisional material)
0.1.0.70–71	Chiang hair BSDF (MDL SDK port, OmniHair inputs); GPU memory log; Visible-off nodes skipped
0.1.0.72	Strand hair as procedural curve segments (4x faster, hair memory ~2.3 -> 0.4 GB)
0.1.0.73	PBRSkin makeup layer
0.1.0.76	Render subdivision of cage meshes (SDK Catmull-Clark of a copy; no HD morph detail)
0.1.0.77–78	Render failure cleanup; shader specialization and index medium stack (1.3–3.3x on benchmarks)
0.1.0.80	Iray Uber Dual Lobe Specular Reflectivity map (living-room skin 1.21 -> 1.00)
0.1.0.81–85	Daz camera headlamp as Iray renders it (render_test4 body view 0.81 -> 1.01)
0.1.0.86	Full-resolution unfiltered textures for metallic materials with normal/roughness maps (glitter top 0.80 -> 1.00)
0.1.0.87	Learned path guiding and emitter-guided paths removed (no host gain)
0.1.0.88	Material self-reporting ("Not translated [...]" log lines); PBRSkin Bump Enable honoured
0.1.0.89	First-hit material map beside the PFM; per-material host-check regions
0.1.0.90	Canonical material layout in the shader (M1 step 1; images bit-identical)
0.1.0.91	Canonical offline::Material in core, plugin and tests (M1 step 2a; bit-identical)
0.1.0.92	Translators write the canonical material; former layout deleted (exported materials byte-identical)
0.1.0.93	Material translators split per shader family on ChannelReader (byte-identical)
0.1.0.94	Translator golden tests from captured material snapshots
0.1.0.95	Iray Uber top-coat reflectivity read (was always the default); M1 complete
0.1.0.96	Makeup maps in texture footprints; OmniHair opacity per its MDL
0.1.0.97	PBRSkin second specular lobe, top coat and detail maps
0.1.0.98	Diffuse roughness (libbsdf Oren-Nayar), Iray Uber and PBRSkin (probe-verified)
0.1.0.99	Iray Uber Diffuse Overlay (probe-verified)
0.1.1.0	Iray Uber Backscattering, libbsdf lobe ported (probe-verified)
0.1.1.1	Thin-walled dielectrics as one Fresnel event (eye-moisture catchlights)
0.1.1.4	Metal-detail texture rule only for metallic maps with metal content
0.1.1.5	libbsdf diffuse normalization for tilted shading normals (normal-mapped surfaces)
0.1.1.6	MDL anisotropy conversion without a cap (card hair)
0.1.1.7	Camera headlamp scales with Focal Distance² as in Iray (large rooms)
0.1.1.8	Iray's Gaussian pixel filter (radius 1.5 px); Crush Blacks tone-mapping setting
0.1.1.10	Depth of field, aperture blades, orthographic cameras, crop windows; vertical field of view (non-square framing fixed)
0.1.1.13	Iray Sun-Sky (base::sun_and_sky port, Daz's solar position)
0.1.1.17	Environment and sun light through flat glass panes (thick and zero-thickness) with MIS
0.1.1.20	Daz spot lights (Iray cone profile and intensity), photometric lights (flux, colour temperature)
0.1.1.21	Code-review fixes: spot lights selected by flux; texture footprints for orthographic and lens-shifted cameras
0.1.1.22	Photometric lights honour Intensity; a dome without a map renders Iray's Sun-Sky
0.1.1.23	Iray Environment Tint on dome and Sun-Sky environments
0.1.1.24	Daz instance nodes (flattened copies, nested instances, instance groups)
0.1.1.26	Iray Matte Fog (distance fog on every path segment, in-scatter relative to the environment)
0.1.1.29	Daz area lights: photometric rectangle, disc and sphere lights with the spot profile, visible in reflections
0.1.1.30	Code-review fixes: mirrored instance copies keep outward winding; instance transforms read exactly
0.1.1.34-0.1.1.38	Previews on a worker thread, preview denoise interval, path diagnostics off by default, texture cache without copies
0.1.1.41-0.1.1.42	Sun-Sky follows Dome Rotation; dome tilt (Dome Orientation X/Y/Z)
0.1.1.47	Iray ground plane / shadow catcher (Draw Ground)
0.1.1.51	Geografts (the follow target's assembly mesh)
0.1.1.52	Default bounce limit 32
0.1.1.54	Window sheet pairs crossed as one pane
0.1.1.56	Matte Fog under map domes
0.1.1.57	Ground-plane sun shadows
0.1.1.58	EXR / floating-point textures
0.1.1.60	Window sheet pairs at oblique angles, passed as two thin-walled sheets
0.1.1.61	White Point tone-mapping setting
0.1.1.62	Dome Only / Sun-Sky Only drop Daz light nodes, as in Iray
0.1.1.63	Base-colour texture alpha no longer used as cutout
0.1.1.71	Faster render setup (re-render of the street scene 4.77 -> 3.43 s)
0.1.1.72-74	Shader displacement (Displacement Strength map, Min/Max, SubD Displacement Level; probe silhouettes equal to Iray's)
0.1.1.75-77	Deterministic renders (learned light selection in fixed point); persistent GPU context and on-disk pipeline cache
0.1.1.78	Image series / animation renders; renderer core split into prepare / sample / diagnostics phases
0.1.1.79	Scene exporter split into per-family files (src/daz/export/), renders unchanged
0.1.1.80	Cutout coverage independent of the GPU's traversal order (run-to-run exact unless Daz's input varies)
0.1.1.81	Graded Ghost Light Factor (glossy highlights fade as in Iray); coloured Glossy Specular per channel
0.1.1.82	Colour images on scalar channels read as Iray does (RGB average); glass roughness maps applied
0.1.1.83	Glossy materials keep full-resolution normal maps (highlight glints as in Iray)
0.1.1.84	Geometry export 1.7-3.2x faster (identical output)
0.1.1.85	Glossy Reflectivity maps
0.1.1.86	Deterministic ray-hit tie-break
0.1.1.87	Glossy materials with normal maps keep all maps at full resolution (brick and stucco walls as in Iray)
0.1.1.88	Repeat renders skip the closed-volume check (session cache)
0.1.1.91	Iray ground plane: only sky, sun and directional lights cast its shadow (sunlit interiors as in Iray)

Project map

• src/daz: plugin registration, native render properties, render lifecycle, GUI-thread scene export and progressive Daz handler output.
• src/core: value-only scene data, materials, lighting preparation and Vulkan resource/BLAS/TLAS ownership. GPU workers never access live Daz scene objects.
• shaders/pathtrace.comp: the path tracer (path(), main()), including shaders/pathtrace in dependency order: texture filtering, strand-curve intersection, layered surface BSDFs (diffuse, GGX, Iray simple_glossy, top coat, dual lobe), the Chiang hair BSDF, dielectrics and media, subsurface walks, emission, direct lighting and light selection, refractive connections and MIS (layout: ARCHITECTURE.md).
• tests: CPU invariants, native SDK/runtime checks and synthetic GPU fixtures. These do not replace testing inside Daz.
• tools: build and user-operated installation helpers; tools/analysis: Iray comparison and asset-inspection scripts, and harness suites in tools/analysis/suites; tools/probes: headless DAZ Script helpers (scene inspection: environment, lights, instances, tone mapping, camera, hair) and probe scenes (probe_metal_point.duf, probe_lights.duf, render_test4 with Iray canvases).

Documents:

• SCOPE.md: goal, fidelity bar, in/out of scope, roadmap and the user's decision log.
• HANDOFF.md: current status and next steps.
• AGENTS.md: how agents work on this project (reading order, working style, documents, build, release, host check).
• ARCHITECTURE.md: data flow, modules, invariants, debt, the target material model and the techniques as implemented.
• COVERAGE.md: which shaders content uses and per channel what is verified, derived, assumed, approximated or missing (with usage counts from the content library); the same for geometry, lights, environment, camera, output, animation and scale, with scope tiers.
• MATERIALS.md: Iray Uber/PBRSkin/OmniHair channel bindings and units.
• LEARNINGS.md: lessons by topic.
• VALIDATION.md: chronological commands, results and caveats from October 19 (0.1.0.74 onward).
• archive: historical only. Validation for 0.1.0.1–0.1.0.18, 0.1.0.19–0.1.0.47, 0.1.0.47–0.1.0.57, 0.1.0.58–0.1.0.65 and 0.1.0.66–0.1.0.73, RELEASE_HISTORY.md (releases through 0.1.0.14) and NOISE_REVIEW.md (an October 1 independent review, mostly addressed).

Build

Prerequisites: Python 3, CMake 3.22+, MSVC 2022 Build Tools with Windows SDK, Vulkan SDK with glslc (local: 1.4.357; the renderer needs Vulkan 1.2 with ray queries and 64-bit buffer atomics), the matching Qt development installation, and the Intel Open Image Denoise 2.5.1 Windows release extracted to ../../.build-tools/oidn/oidn-2.5.1.x64.windows (override with -DOIDN_DIR). Use Release or RelWithDebInfo, never a Qt-debug DLL in Daz.

From this plugin directory:

python .\tools\Build.py --sdk-dir '..\..\Daz Studio 6.25+ BETA SDK' --test

The helper also defaults to that older sibling SDK, never Latest. Qt is located independently, checking the selected SDK, workspace .build-tools and the reference SDK's .build-tools. Override it if needed:

python .\tools\Build.py --sdk-dir '..\..\Daz Studio 6.25+ BETA SDK' --qt-dir 'C:\Qt\6.10.3\msvc2022_64\lib\cmake\Qt6' --test

New builds use build/current (or build-core/current for --core-only). build/releases/<version> holds every released DLL and shader (experiment builds pruned; PDBs from 0.1.1.17; one shared denoiser runtime in build/releases/denoiser-runtime) and build/fixtures the real-asset test fixtures. --build-dir overrides the output folder; a cache from another source location is rejected with a recovery instruction. The helper normalizes Windows environment-key casing before launching tools.

python .\tools\Build.py --core-only --test
python .\tools\Build.py --test --gpu-test

--test runs build-helper regressions, CPU CTest and, for a plugin build, native settings, optics, emission-color and loader checks against the installed runtime. --daz-dir overrides the runtime directory. --gpu-test runs the full synthetic GPU suite serially; it is not a Daz GUI render. Shader validation is recorded separately in VALIDATION.md.

Default DLL: build/current/plugin/RelWithDebInfo/dsp_vulkan_offline.dll.

Install and host testing

Close Daz first. The user's account owns the installed plugin DLL and PDB, so the helper runs without elevation (agents may run it: AGENTS.md); a first install, or a changed denoiser runtime, needs an elevated PowerShell:

.\tools\Install.ps1 -DazDir 'C:\Program Files\DAZ 3D\DAZStudio6'

The helper backs up the existing DLL to build/install-backups (last five kept), overwrites only the plugin DLL/PDB (not Qt libraries), copies the denoiser runtime only when it changed, and prints the installed SHA-256. -Release 0.1.0.73 installs an archived release from build/releases (never 0.1.0.39 or 0.1.0.41); -BuildDir selects a nondefault build folder. Never copy development Qt DLLs into Daz. Restart Daz and select Vulkan Offline (Experimental). Host renders and Iray comparisons run headless through the harness in ../../daz tools (runs and the Iray reference cache live there); GUI renders are for what the harness cannot do (host checks: build/host-checks). Use the host-check instructions in AGENTS.md, not historical pending requests in the archive. Daz's host log is %APPDATA%\DAZ 3D\Studio6\log.txt, not in the installation directory.

Scripting (DAZ Script)

Settings are properties of the renderer's settings element, reachable from the active renderer (0.1.0.54):

var renderer = App.getRenderMgr().getActiveRenderer(); // Vulkan Offline (Experimental)
print(renderer.getRenderSettingNames());
renderer.setRenderSetting("SamplesPerPixel", 128);
renderer.setRenderSetting("Denoiser", "Final image");   // or 0, 1, 2
renderer.setRenderSetting("WriteLinearImage", true);
print(renderer.getRenderSetting("SamplesPerPixel"));

Changes apply to the next render and persist like pane edits (registry HKCU\Software\VulkanOffline\DazStudio6; set the environment variable VULKAN_OFFLINE_SETTINGS_FILE to an INI path to use a separate file). When a render goes to a file (render options DirectToFile with a filename), the linear PFM is written next to it with the same base name.

Features and limitations

What is implemented, approximated or missing: COVERAGE.md; how it works: ARCHITECTURE.md.