This website requires JavaScript.
61c63e513e
feat(mflux_image_edit): pass ALL inputs to --image-paths — multi-image compose
2026-07-20 18:33:19 +1000
3b8c88ce48
fix(sd_local): multi-dir LoRA search + UNet-only load
2026-07-20 18:20:12 +1000
f374465216
docs: MRP_MLX.md — canonical reference for the MLX stack (pillars, perf, knobs, lessons ledger, runbooks, roadmap)
2026-07-20 17:07:56 +1000
de0b509dfa
operator: trellis2_mlx — the 72s/26GB MLX fork as a farm operator (vertex baker default; m1 routed via nodes.json)
2026-07-20 17:04:20 +1000
eb3862bb8a
rung 3 final: M1 Ultra validated — 181.5s / 17.2GB, both Ultras now run optimized MLX TRELLIS.2
2026-07-20 04:16:29 +1000
95aa214b1e
rung 3 results: 3.0x e2e (216.6->72.0s), 2.8x memory, parity held — sampler+baker+validator on the fork
2026-07-20 03:55:54 +1000
d0b02003b8
benchmarks: Qwen-Layered complete ladder — q8 floor (q6/q4 quality cliff), 1024 hero tier, operator shipped; flux_local audit (FLUX.1 compile gap)
2026-07-19 22:20:49 +1000
58a2a8177d
operator: qwen_layered_local — image -> N editable RGBA layers (Qwen-Image-Layered 20B, MLX)
2026-07-19 22:06:05 +1000
8849438bce
benchmarks: Qwen-Layered cross-bench — m3ultra 243.0s (1.50x vs m1), both Ultras serving from baked q8
2026-07-19 21:41:28 +1000
0db2c90481
benchmarks: Qwen-Layered ladder — 20 steps holds quality, 3.2x to 365.6s on m1ultra
2026-07-19 21:09:37 +1000
64afe54474
benchmarks: Qwen-Image-Layered working on MLX (m1ultra) — 4-layer RGBA decomposition verified, PR302 fork vendored
2026-07-19 20:40:02 +1000
79bfd713c1
docs+benchmarks: vertex baker productionized in both pipelines — 130s e2e, 1s bake, zero dark patches
2026-07-19 19:17:01 +1000
232cc84420
bake fix PROVEN: KDTree-IDW vertex sampling kills the dark patches in 0.3s; axis + gamma gotchas documented; production path defined
2026-07-19 18:07:21 +1000
5d2a3f57e4
benchmarks: KDTree control post-mortem — 5.85h bake + broken export; fast-rasterizer rewrite required
2026-07-19 15:52:26 +1000
142a0ac5da
docs: packet — trellis-2-mrp-mlx fork pushed to Gitea (pin tag pr175-head-754d403)
2026-07-19 15:30:39 +1000
603efcf7e2
docs+benchmarks: dark-patch verdict — decoder attrs near-fal clean, Metal texel-sampling convicted; regular-grid gather fix designed; parity fixtures complete
2026-07-19 14:03:53 +1000
a0e1ef1f18
benchmarks: Hunyuan mlx-tune final — 243s vs 260s (1.07x), M1 rebench suggested
2026-07-19 13:54:46 +1000
d488d87490
benchmarks: Hunyuan mlx-tune e2e 246s vs 260s baseline (~1.06x, texture-led)
2026-07-19 12:51:24 +1000
0476d1bc30
docs: packet §9 — Fable session 1 execution log (oracle pass, A0 ADOPT, mlx-tune branch, dark-patch narrowed)
2026-07-19 12:47:11 +1000
0ac173d4fb
benchmarks: A0 gate — MLX TRELLIS.2 (Jourloy hybrid) BEATS MPS 216.6s vs 248.8s at 14x bake density; dark patches reproduce -> Metal texel sampling prime suspect
2026-07-19 12:46:16 +1000
ace4be37ba
benchmarks: Hunyuan T1 mx.compile micro gate — parity PASS, 1.04x (GEMM-bound; T3/T4 are the real levers)
2026-07-19 12:40:06 +1000
eed732446c
docs: FABLE_STARTUP_PACKET §8 — adopt path needs NO weight conversion (Jourloy loads originals directly; convert_weights.py retired as double-transpose trap)
2026-07-19 11:55:38 +1000
3b6b9124f9
docs: FABLE_STARTUP_PACKET §7 — round-2 verification (ADOPT confirmed HIGH, transpose policy verified, parity oracle CPU-runnable, BiRefNet fetched; flags ss_dec double-transpose + bf16-crash + downcast in convert_weights)
2026-07-19 11:51:47 +1000
abfa344565
docs: FABLE_STARTUP_PACKET — prep recon consolidated (adopt PR#175 MLX base, Hunyuan mx.compile targets, prep_local draft, weight-convert+parity plan)
2026-07-19 11:11:23 +1000
ca4c802e40
docs: master build plan — TRELLIS.2+Hunyuan full MLX + kernel rewrites + preprocessing frontend (multi-day roadmap for Fable)
2026-07-19 10:46:38 +1000
cf402e322c
fix(stereo_depth): negative disparity window + percentile normalize — movie stereo works
2026-07-19 01:51:55 +1000
95708ad220
feat: sd_local (SD1.5 + anatomy LoRAs + pose/IP-Adapter) + stereo_depth (SBS → SGBM depth/cloud)
2026-07-19 01:22:12 +1000
fcb5e80702
benchmarks: trellis_mac sparse-conv ablation - Metal flex_gemm = 1.72x end-to-end vs pure-torch fallback
2026-07-19 00:50:56 +1000
793f8a88bb
benchmarks: m1max studio added to fleet MLX matrix - 7.1 TF, 336 GB/s second-best bandwidth
2026-07-19 00:35:15 +1000
5bd7191b32
scripts: mlxprobe.py — fleet MLX capability probe
2026-07-19 00:31:59 +1000
ed2ee8dc16
benchmarks: 6-machine fleet MLX matrix — bf16-on-M1 answered, head_dim-64 padding critical on M4/M5
2026-07-19 00:31:55 +1000
daa6124dcf
benchmarks: first trellis_mac vs fal trellis-2 head-to-head — local faster, quality gap = hole-fill + pre-simplify
2026-07-19 00:22:41 +1000
2734ef9107
docs: TRELLIS.2 Apple Silicon recon — trellis-mac is already TRELLIS.2, MLX ladder, PR #175 watch
2026-07-18 23:44:30 +1000
d5217575eb
wip(corridorkey_local): operator scaffold — BLOCKED on GVM inference crash
2026-07-18 22:30:39 +1000
4bc2f56d34
README: local-first rewrite + ignore the .env.* family
2026-07-17 22:24:18 +1000
b9de2a4167
docs: register the M1 Max / M4 mini / M1 mini nodes in CLUSTER + HARDWARE
2026-07-17 22:24:36 +1000
712a6ed520
install_acestep: torchcodec for torchaudio>=2.9 save path
2026-07-17 19:47:03 +1000
12a8cf1fd2
music_local (ACE-Step) + motion_local (MoMask) operators with install scripts
2026-07-17 19:42:14 +1000
b7e113ca5b
wan/lipsync: scratch dirs in /tmp so they don't register as assets
2026-07-17 18:57:54 +1000
8ea7e8928b
stt/wan run.py: prepend homebrew to PATH (job env can't find ffmpeg)
2026-07-17 18:42:43 +1000
110b2298b7
audio/video/text operator suite: stt_local, lipsync_local, voice_clone_local, wan_video, llm_local
2026-07-17 18:26:32 +1000
27564e80eb
perfcheck: nightly fleet GPU canary + the contention lessons that shaped it
2026-07-17 16:50:06 +1000
f6626f0a06
tts_local operator: text→speech via Kokoro-82M on MLX (character voice templates)
2026-07-17 18:01:58 +1000
26c4cfd099
benchmarks: 8-machine CorridorKey matrix — tiled keys 2048 on the 2020 M1 mini
2026-07-17 15:04:00 +1000
a6a6d110b2
benchmarks: NEGATIVE result — I2V-synthesized orbits don't reconstruct in 3D
2026-07-17 02:07:19 +1000
acec2cc6c5
benchmarks: first M3 Ultra mlx-video/Wan numbers + the auto-tiling SIGKILL gotcha
2026-07-17 01:47:48 +1000
9c7164f77a
comfyui_sd: fail fast when a checkpoint/LoRA isn't on the routed node
2026-07-17 01:44:17 +1000
75ace36e00
comfyui_sd operator: local SD/SDXL + LoRA via resident ComfyUI
2026-07-16 23:37:30 +1000
a0f4114c28
scripts/install_comfyui.sh: portable ComfyUI install (SD/SDXL runner; mflux is FLUX-only)
2026-07-16 23:28:25 +1000
fe0da76016
benchmarks: fleet ablation results — sdpa 4.5x M1 cliff (bf16 innocent), tiled768+compile new best, M4 swap wall
2026-07-16 22:48:50 +1000
532294b7a5
users.py: 'token <name>' subcommand — mint named API keys from the CLI
2026-07-16 22:29:08 +1000
273922f492
CORRIDORKEY.md: status, Ultra tuning findings, video->3D use-case map
2026-07-16 21:50:56 +1000
175e1eeac4
benchmarks: CorridorKey MLX Ultra tuning — tiled512+compile patch wins (1.47x M3, best quality, 2.3GB)
2026-07-16 21:38:08 +1000
876a666e9a
HARDWARE.md: per-operator RAM + Apple Silicon tier guide (8GB→256GB)
2026-07-16 14:00:39 +1000
1537d94cf7
benchmarks: 4096 hunyuan defaults confirmed fleet-safe (M3 + M1)
2026-07-16 11:11:33 +1000
06f33be865
pool cpu lane across nodes; hunyuan Studio-quality defaults
2026-07-16 10:59:44 +1000
9b525fa420
hunyuan install clones self-owned MLX fork; CLUSTER fleet final state
2026-07-16 10:35:11 +1000
573afa7b7a
hunyuan3d_mlx operator (local MLX image→3D) + cluster docs
2026-07-16 10:10:43 +1000
19cf9ed21e
🔑 agents get keys: shared agent token minting + guest-visible job outputs
2026-07-15 23:30:04 +1000
b8c333a93e
Unified worker pool: gpu lane dispatches across M3 + M1 nodes (HANDOFF2 phase B)
2026-07-14 18:18:09 +1000
5f46939321
runner: gpu lane becomes a node pool (local + remote workers)
2026-07-14 13:38:34 +1000
d8d44bad45
docs: M1 Ultra second node (100.91.239.7) in AGENTS.md
2026-07-14 12:01:22 +1000
a4facbacd9
serve.sh: mkdir -p data before logging (fresh-checkout fix for new nodes)
2026-07-14 11:51:29 +1000
210d80484c
runner: resolve relative operator python paths against repo root (multi-node portability)
2026-07-14 11:47:39 +1000
37cf85693a
docs: auth + dashboard + public hosting (README, AGENTS.md)
2026-07-13 12:31:27 +1000
514ec6cdcc
Security hardening from adversarial review (5-agent workflow)
2026-07-13 12:25:13 +1000
810400fb60
Auth + dashboard (HANDOFF2 phases 1-2): guests local-only, system snapshot
2026-07-13 12:06:52 +1000
4e66cc632e
HANDOFF2: execution brief — multi-user auth (guests local-only), dashboard, VPS hosting
2026-07-13 11:10:27 +1000
8fc7b407f3
docs: CUDA worker — add Win11 + WSL2 path (recommended, no reinstall)
2026-07-13 11:01:36 +1000
27ce69782d
docs: CUDA worker setup guide (T490s + 2080Ti eGPU on tailnet)
2026-07-13 10:58:32 +1000
4b73b90625
BENCHMARKS: TRELLIS.2 local working — 289s/~4.8min gen, SOTA quality
2026-07-13 10:23:37 +1000
95954a62d4
AGENTS.md: real fal prices from owner billing (trellis2 $0.05, hunyuan mv $0.017); fal+openrouter keys set
2026-07-13 09:50:02 +1000
764ae3faf5
trellis_mac: unblock — dinov3 lvd1689m granted; add HF_HUB_DISABLE_XET + OMP guards
2026-07-13 09:44:13 +1000
f0bac67edb
Local mflux operators: seedvr2_upscale + mflux_image_edit (free, ungated)
2026-07-13 01:37:56 +1000
1cc0b80d89
bg_remove_local: local RMBG-2.0 background removal on MPS
2026-07-13 01:32:38 +1000
21938a4e17
sf3d: fix macOS OpenMP segfault — single-threaded OMP + KMP_DUPLICATE_LIB_OK
2026-07-13 01:16:40 +1000
c42f066723
flux_local: HF_HUB_DISABLE_XET fix + full local FLUX lineup benchmarks
2026-07-13 00:37:24 +1000
6c54091130
BENCHMARKS: first local-vs-cloud image A/B (Klein 4B vs nano-banana)
2026-07-12 23:27:37 +1000
7c0914f714
BENCHMARKS: first M3 Ultra FLUX.2 Klein 4B numbers — 8s/image at 1024px, 18GB peak
2026-07-12 22:42:41 +1000
19f086773e
Headless control: mb CLI, AGENTS.md handover, serve scripts; verified-current image models
2026-07-12 22:38:52 +1000
76f065d3a6
A/B image-gen: local FLUX (mflux/MLX) + OpenRouter nano-banana operators
2026-07-12 22:31:19 +1000
7ea3c8b935
fal panel: 5 new operators, image outputs, grouped UI, no-input ops + review fixes
2026-07-12 21:51:16 +1000
605b1ae347
Phase 1 + framework: settings/secrets, queue lanes, job mgmt, inbox, 8 new operators
2026-07-12 21:42:27 +1000
d0569dc746
Phase 0 baseline: FastAPI+SQLite operator pipeline, React/three.js UI, ffprobe/ffmpeg_frames/blender_convert operators
2026-07-12 21:05:22 +1000
type-two