benchmarks.csv

workflow,model,input_spec,output_spec,wall_clock_sec,peak_vram_gb,throughput,concurrency,marginal_cost,cloud_equivalent_cost,date,notes,sample_output
Flashfire,flux-2-klein-4b-fp8 (distilled),"prompt-only, 4 steps, CFG 1.0, seed 42",1024x1024 PNG,10,,10s/image,n,SGD 0.001,USD 0.010,,render-verified 2026-06-15 (cold-load incl); FLUX.2 4B distilled,samples/flashfire/output_00003_.png
Goldsmith,flux-2-klein-base-4b-fp8,prompt-only 20 steps,1024x1024 PNG,6,11.2,6s/image,1,SGD 0.001,USD 0.012,2026-06-13,sweep: confirms ~3.3s ballpark (6s measured),samples/goldsmith/output_00004_.png
Silkmotion,RIFE v4.9 (frame-interpolation),video in,60fps out,12,3.5,12s/clip (RIFE interp),1,SGD 0.002,USD 0.040,2026-06-13,sweep: confirms ~8.3s ballpark,samples/silkmotion/output_00001.mp4
Crystalforge,seedvr2_ema_7b_fp16 (v2.5),"672x384 24fps 6.0s (VOID output), batch 5, res 1080",1890x1080 24fps 6.0s H.264,586,25.7,586s/clip,1,SGD 0.077,USD 0.200,2026-06-12,"E1b chain stage 2: clean upscale, no artifacts at inspected crops; VOID->Crystalforge total ~13 min for 6s deliverable",samples/crystalforge/e1b_00001.mp4
Vanisher,VOID pass1+pass2 + SAM3 + cogvideox_vae,"1366x768 24fps 5.9s clip, SAM3 person mask",672x384 24fps 5.0s (HALF-RES + truncated — see notes),149,26.8,149s/clip,1,SGD 0.020,USD 0.500,2026-06-12,"E1 bench: mid-clip removal clean (fill-region frame-delta 7.0-8.0 vs 8.3 control = stable); FAIL: partial-removal ghost at t=0.5s where subject crouched/occluded; output half-res + 0.9s truncated -> needs Crystalforge pass + frame-cap fix before client delivery || E1b (tuned: 6s cap, lab fill prompt, SAM3 thresh 0.35): 189s / 26.9GB, FULL 6.0s duration; ghost reduced but present — root cause: image-mode SAM3_Detect lacks temporal propagation (E1c: port SAM3 video pipeline) || E1c (SAM3 IMAGE->VIDEO pipeline grafted: LoadSAM3Model+VideoSegmentation+Propagate+VideoOutput replacing SAM3_Detect; fixed comfy-env==0.3.89 missing dep that blocked the video nodes): 280s/28GB, validated+runs, masks now temporally propagate. Ghost REDUCED but not eliminated on this worst-case clip (subject behind glassware, low light) — residual is now a VOID inpainting-quality limit, not a mask-tracking failure. Recommend re-test on a normal-lit clip with a cleanly-separated subject. || E1d (representative clip: well-lit kitchen, cleanly-separable spoon, SAM3-video pipeline from E1c): 180s/27GB — CLEAN PASS. Spoon removed, NO ghost, NO temporal smear, background (steel equipment + counter) reconstructed perfectly. Confirms thesis: the E1 ghost was a worst-case-clip artifact (subject behind glassware in low light), not a pipeline failure. Client-deliverable on representative footage.",samples/vanisher/void-e1d-kitchen_00001_.mp4
Shapeshifter,wan2.1_vace_14B_fp16 + SAM3 video,video + SAM3 video-mask + change prompt,masked region edited,802,20.6,802s/clip (VACE edit),1,SGD 0.106,USD 0.600,2026-06-13,sweep: VACE+SAM3-video runs clean,samples/shapeshifter/output_00001.mp4
Wan22_I2V,wan2.2 14B fp8 (high+low MoE),keyframe + motion prompt two-sampler,720p 81fr MP4,502,20.1,6.2s/frame (149x realtime),1,SGD 0.066,USD 0.350,2026-06-13,sweep: confirms ~506s est; keyframe animation verified,samples/wan22-i2v/regression_seed42_20260528.mp4
Wan22_T2V,wan2.2 14B fp8 (high+low MoE),prompt-only two-sampler seed 42,720p 81fr MP4,483,20.0,6.0s/frame (143x realtime),1,SGD 0.064,USD 0.300,2026-06-13,sweep: confirms ~497s catalogue est; two-sampler MoE verified,samples/wan22-t2v/output_00001.mp4
Hunyuan_I2V,hunyuanvideo 1.5 (distilled),"keyframe + prompt, CFG 1.0, flow_shift 5.0",video MP4,21,,"21s/clip (=quickening,5s out)",n,SGD 0.003,USD 0.300,,superseded by quickening; verified 2026-06-15,
Hunyuan_T2V,hunyuanvideo 1.5,prompt-only,video MP4,16,,"16s/clip (=dreamforge,5s out)",n,SGD 0.002,USD 0.250,,superseded by dreamforge; verified 2026-06-15,
WanDraft,wan2.1 1.3B,prompt-only draft,low-res preview MP4,70,,70s/clip,n,SGD 0.009,USD 0.080,,render-verified 2026-06-15; Wan2.1 1.3B draft video; SFW sample seed 42 rendered 2026-07-02,samples/quickdraw/quickdraw_00001.mp4
FluxCommercial,flux1-dev,prompt-only fast draft,high-res PNG,20,,20s/image,n,SGD 0.003,USD 0.025,,"render-verified 2026-06-15; FLUX.1-dev clean product shot (prev. broken, now rebuilt). LICENCE: FLUX.1 [dev] Non-Commercial License — NON-COMMERCIAL, fast drafts only, NOT for paid/client work (corrected 2026-06-22; was mislabelled commercial-safe). Qwen-Image (Apache-2.0) is the commercial-safe keyframe hero and the --mode commercial default (2026-06-22); Chroma1-HD (Apache-2.0) stays as the artistic alt.",samples/fluxcommercial/output_00001_.png
QwenImage,qwen-image-fp8-e4m3fn (20B Apache-2.0),"prompt-only, 20 steps, CFG 2.5, ModelSamplingAuraFlow shift 3.1, seed 42",1024x576 PNG,12,28,12s/image warm (~16-20s cold-load),1,SGD 0.002,USD 0.020,2026-06-22,"MEASURED 2026-06-22: Qwen-Image commercial-safe HERO (Apache-2.0 — genuinely OK for paid/client work, unlike FLUX.1-dev/ledger which is Non-Commercial). ComfyUI native graph: UNETLoader(qwen_image_fp8) + CLIPLoader(qwen_2.5_vl_7b, type=qwen_image) + VAELoader(qwen_image_vae) + ModelSamplingAuraFlow shift 3.1 + KSampler euler/simple 20 steps CFG 2.5. Wall-clock 12s warm / 16-20s cold (incl. ~30GB model load), peak ~28GB VRAM (DiT 20.4GB fp8 + text-encoder 9.4GB; fits 32GB w/ ~4GB headroom). Eye-verified across 3 seeds (42/99/123): clean commercial product shots, strong prompt-adherence, zero artifacts. WIRED as --keyframe qwen + promoted to the --mode commercial DEFAULT (was chroma). Strengths: prompt-adherence + native text rendering. GPU etiquette: ran in free window, ComfyUI /free between measurements, evicted no foreign proc.",samples/qwenimage/commercial_keyframe_00003_.png
DailyPipeline,Dolphin/LLM -> FLUX -> Wan22 I2V,text brief -> keyframe -> video,keyframe PNG + final MP4,510-526,,"510-526s end-to-end (keyframe + i2v, no score)",y,SGD 0.067,USD 0.350,2026-06-22,"MEASURED 2026-06-22 (was estimate): derived from the maestro + commercial full-pipeline A+B components measured today — keyframe 8-16s + Wan2.2 I2V MoE ~510s, no music score. Bounds the maestro 658s full pipeline minus the ~8s ACE-Step score and minus the keyframe spread (klein 8s .. ledger 16s).",samples/daybreak/final_video_00001.mp4
VideoWithMusic,FLUX -> Wan22 I2V -> ACE-Step,text brief -> video + music,final MP4 w/ audio,787,,787s end-to-end,y,SGD 0.104,USD 0.600,,= maestro realized 2026-06-15,
Tier6_Finishing,RIFE -> SeedVR2,30fps 720p clip,60fps 2160p clip,~600,,~600s (est: Crystalforge 4K 586s + Silkmotion 60fps 12s),n,SGD 0.079,USD 0.400,,estimate from measured components,
dreamforge,hunyuanvideo1.5_720p_t2v_fp16,"T2V prompt-only, 512x512, 25 frames",512x512 mp4,16,,16s/clip (5s out),n,SGD 0.002,USD 0.250,2026-06-15,render-verified — paper airplane over city,samples/dreamforge/dreamforge_f0.png
quickening,hunyuanvideo1.5_720p_i2v_cfg_distilled_fp8,"I2V from image, 512x512, 25 frames",512x512 mp4,21,,21s/clip (5s out),n,SGD 0.003,USD 0.300,2026-06-15,render-verified — image animated,samples/quickening/quickening_f0.png
sharpscale,hunyuanvideo1.5_1080p_sr_distilled_fp8,384 base → 1080p super-res,1920x1072 mp4,73,,73s/clip (5s out),n,SGD 0.010,USD 0.120,2026-06-15,render-verified — true 1080p upscale,samples/sharpscale/sharpscale_f8.png
maestro,PRISM(:8001) LLM -> FLUX.2-Klein-4b -> Wan2.2 I2V MoE -> ACE-Step,scene idea (one line text),832x480 24fps MP4 + 30s original score,512,18,512s end-to-end (kf+i2v+ACE score),n,SGD 0.067,USD 0.600,2026-06-15,"render-verified: sharp lighthouse keyframe (1024x576; Klein melts >1MP/720p), clean I2V motion, real ACE-Step instrumental score muxed. kf 4s + i2v 500s + ACE-Step music 8s",samples/maestro/maestro_s42_final.mp4
Inkwell,Dolphin 3.0 R1 Mistral 24B AWQ (awq_marlin INT4),"logline/premise, temp 0.07, single-stream",long-form screenplay/script text,~105 (cold-load to healthy),28.6,106 tok/s single-stream decode (median of 5 runs 512-1024 tok; min 106.2 max 106.6),1,SGD <0.001,USD ~0.10/1M out,2026-06-22,"MEASURED (was ~80 tok/s est): GPU-resident, NO cpu-offload, gpu-mem-util 0.80, max-model-len 32768, FLASHINFER, FA2. DECODE TRAP FIXED: shipped tokenizer_config declares tokenizer_class=LlamaTokenizer -> transformers 5.x routes to broken Mistral-regex slow path -> raw byte-level BPE leak (G-dot/C-dot for space/newline, the queue.md C-trap ~2 tok/s+garbage symptom). Fix: override tokenizer_class=PreTrainedTokenizerFast in a tokenizer-only overlay dir (weights stay read-only), serve via --tokenizer; decode then 0 BPE artifacts. KV cache 63648 tok (32K ctx usable, 1.94x concurrency). 450W sustained, 58C, no throttle. Co-resident w/ idle ComfyUI+audio (~3.1GB foreign).",
FaceForward,klein(FLUX.2-4b) / lustify(SDXL) / chroma(Chroma1-HD),"same scene+seed42, --face-forward A/B (stage A keyframe)",1024x576 PNG keyframe (lustify 1216x832),kf 4-20s + I2V 140s (draft 25fr),29.4,klein ~4s / lustify ~4-8s / chroma ~12-20s per keyframe,n,SGD ~0.001,USD ~0.012,2026-06-22,"CRE-4 VERIFIED. (1) KEYFRAME A/B: --face-forward flips subject profile->frontal on all 3 supporting keyframe models, same seed 42. klein = POSITIVE-ONLY (cfg 1.0, empty neg) -> orientation goes frontal but gaze can stay ambiguous under backlight; lustify+chroma = POS+NEG (FACE_FORWARD_NEG suppresses profile/turned-away) -> clean frontal portrait w/ direct eye contact. chroma at C-17 full steps (26 / CFG 4.0). (2) I2V-SURVIVAL (BS-9) VERIFIED: face-forward klein kf -> Wan2.2 I2V MoE 30-step -> maestro_video_00002.mp4 (832x480, 25fr draft, 140s); frame0 vs last frame both squared-to-camera, frontal framing SURVIVES animation (no rotate-to-profile). Proof: outputs/cinematic/cre4_{klein,chroma}_ff-{OFF,ON}_seed42.png + cre4_i2v_ff-ON_klein_seed42.mp4 + cre4_i2v_frame_{first,last}.png (lustify pair PRIVATE, local only). PRISM CPU expand_scene pre-step dominated wall time (minutes), not GPU. GPU etiquette honoured: no foreign render evicted (creative-comfyui + Fish-Speech audio co-resident, untouched).",samples/faceforward/cre4_i2v_ff-ON_klein_seed42.mp4
KeyframeAB_ChromaVsLustify,chroma(Chroma1-HD FLUX-based) vs lustify(SDXL),"same SFW scene+seed1234+motion, vary only checkpoint; keyframe + 25fr Wan2.2 I2V draft each",1024x576 / 1216x832 PNG keyframe + 832x480 25fr MP4,"kf: chroma 16s / lustify 4-8s; I2V 120s each",17.8,chroma 17.8GB-load/16s vs lustify 6.9GB-load/4-8s per keyframe,n,SGD ~0.001,USD ~0.012,2026-06-22,"CRE-5 VERIFIED (controlled A/B, C-17 honoured: chroma un-distilled -> 26 steps/CFG 4.0, lustify SDXL 30/CFG 6.0, identical anatomy negative). SFW potter scene chosen to stress hands+face (the structures distilled models melt). FINDING (BS-12 confirmed): LUSTIFY wins human anatomy -> crisp two-hand articulation + sharp face, AND that fidelity SURVIVES Wan I2V better (sharp keyframe gives the animator more structure; BS-9 -> Wan animates softness as faithfully as sharpness). Chroma = softer/painterly/editorial look, lower hand-detail, hands degrade to a blob by last frame under I2V, at 2.6x VRAM + ~2-4x keyframe time. DEFAULT PER JOB: photoreal/anatomy-critical/commercial-realism -> lustify; soft/artistic/editorial/mood -> chroma; fast SFW draft -> klein. Resolves the 'rebuild on Chroma or retire' drift note + handoff 'Chroma never A/B'd': Chroma KEPT, scoped to soft/artistic niche, not the realism default. LUSTIFY's separate PRIVATE path keeps its own numbers in business/ (never here; disclosure is an owner decision, out of scope). Proof: outputs/cinematic/cre5_{chroma,lustify}_keyframe_seed1234.png + cre5_{chroma,lustify}_i2v_seed1234.mp4 + cre5_*_i2v_frame_{first,last}.png. GPU etiquette: ran in a free window alongside idle foreign ComfyUI/audio (~3.2GB), evicted nothing; ComfyUI auto-unloaded between renders.",
CRE-2_CommercialPipeline,FLUX.1-dev 'ledger' keyframe -> Wan2.2 I2V MoE -> ACE-Step -> mux,scene idea (one line) -> commercial full pipeline,832x480 24fps MP4 (3.375s) + score muxed,658,,658s end-to-end (keyframe 16s + i2v 510s + music 8s),n,SGD 0.087,USD 0.600,2026-06-22,"CRE-2 MEASURED 2026-06-22: commercial full-pipeline keyframe->I2V->score->mux. ledger keyframe 16s + Wan2.2 I2V MoE 510s + ACE-Step score 8s = 658s. Output commercial_s42_final.mp4 (3.375s, the Wan I2V single-segment frame cap). LICENCE: the 'ledger'/FLUX.1-dev keyframe is FLUX.1 [dev] NON-COMMERCIAL — fast drafts only, NOT for paid/client work (this row demonstrates the pipeline shape, not a commercially-licensed deliverable). The commercial-safe keyframe is now Qwen-Image (Apache-2.0), the --mode commercial default in cinematic-orchestrator.py as of 2026-06-22 (Chroma1-HD remains the artistic alt).",samples/cre2-commercial/commercial_s42_final.mp4
CRE-3_LongClipChain,FLUX 'maestro' keyframe -> 3x Wan2.2 I2V MoE chain -> ACE-Step -> mux,scene idea (one line) -> long clip via 3-segment chain,"832x480 24fps MP4 (10.125s, 243fr @24fps)",1592,,"1592s total (keyframe 8s + i2v 1480s for 3 segments + music 8s)",n,SGD 0.210,USD 1.500,2026-06-22,"CRE-3 MEASURED 2026-06-22: long-clip via 3x3.375s CHAIN (maestro mode). keyframe 8s + i2v 1480s (3 segments) + music 8s = 1592s. Output maestro_chained.mp4 (10.125s, 243 frames @24fps). DRIFT VERDICT (honest, incl. failures): chaining produces a clean concat BUT inter-segment drift is real + VISIBLE — color palette wanders (blue->pink->blue across segments), the subject's scale/position/shape shifts at each seam, the dolly motion restarts (cadence hitch) at joins, and a seam artifact appeared (malformed gull). Viable for gentle/ambient shots >3.375s but NOT seamless; best <=2 segments, or loop-mode for static scenes. This is the honest data behind the 3.375s single-segment cap caveat.",samples/cre3-longclip/maestro_chained.mp4

coding-stack (AI-1),Qwen3-Coder-30B-A3B-Instruct AWQ-INT4 (vLLM 0.20.0) primary / Qwen2.5-Coder-7B Q5_K_M llama.cpp (:8002) CPU fallback,"coding request via LiteLLM :4000 model=coding, temp 0, single-stream",code/tool-call completion,0.54,28.7,~110 tok/s single-stream decode (GPU primary; 30B-A3B = 3B active MoE),1,SGD <0.001,USD ~0.10/1M out,2026-06-22,"AI-1 BUILT + failover-VERIFIED. GPU primary Qwen3-Coder-30B-AWQ on vLLM :8000 (--moe-backend triton, --dtype float16 [AWQ requires fp16], kv-cache fp8, max-model-len 114688, gpu-mem-util 0.88 -> ~28.7GB, image pinned v0.20.0 proven SM120/Blackwell). Returns proper structured tool_calls (qwen3_coder parser) - VERIFIED. CPU fallback Qwen2.5-Coder-7B GGUF on llama.cpp :8002 (always-on systemd user unit). LiteLLM coding -> coding-cpu failover load-bearing: GPU up=GPU answers; docker stop vllm-coding -> coding fails over to CPU (plain completions OK); GPU restart -> returns to GPU. CAVEAT: CPU tier tool calls are NOT structured (llama.cpp b9283 peg-native does not promote Qwen2.5-Coder tool output to tool_calls; emits text-wrapped call) - GPU is the daily driver, CPU is a degraded keep-typing-during-render backstop. FIREWALL FIX: litellm-proxy attached to docker0 (connect-docker0.sh) so it can reach host CPU endpoints :8001/:8002 past DOCKER-USER (also un-broke pre-existing personal-chain-cpu). ComfyUI VRAM-handoff hook = documented GAP (failure-based fallback makes it correct without it). GPU etiquette: foreign ComfyUI+Fish-Speech (3.25GB) left running, evicted nothing. NOT pointed at Claude Code yet (owner step).",n/a
Scribe-sustained (B-2),Nemotron 3-nano-omni-30b-a3b-reasoning (vLLM 0.20.0 agent mode),"2 concurrent streams (--max-num-seqs 2 cap), 512-tok outputs, 180s sustained window","decode throughput + thermal/power envelope",180.5,27.9,437 tok/s aggregate / 218 per-stream (sustained 2-stream),2,SGD <0.001,USD ~0.10/1M out,2026-06-22,"B-2 clean sustained-load stress (AI-6, end-of-session). Agent-mode Nemotron :8000 (32K ctx, gpu-mem-util 0.82, kv-cache fp8, --max-num-seqs 2). 180.5s steady window (15s warmup discarded): 154 reqs, 78848 tok decoded. AGG 436.9 tok/s, per-stream median 217.8 (215-221), req latency median 2.4s. GPU 96.5% util avg; POWER avg 366.8W / peak 379.8W (well under 450W cap - 3B-active MoE decode is compute-bound on a sparse path, not power-bound); TEMP avg 61C / peak 66C; ZERO throttle across the window; VRAM flat 27.9GB. Replaces the passive ~139-avg observation. Single-stream peak stays 276 (B-1 2026-06-13); knee@8 concurrent=728 agg measured on the forensic config. GPU etiquette: ran in the owner-freed exclusive window; one small foreign proc (2.2GB) left untouched.",n/a

agent-mode shakedown (AI-5b),Nemotron 3-nano-omni-30b-a3b-reasoning (vLLM 0.20.0 agent-light) + Fish Speech (creative-tts),agent-mode.sh from a clean card; concurrent 3x vLLM chat + 1x Fish TTS job via gateway :9000,vLLM :8000 (32K ctx) + audio gateway :9000 coexisting; TTS WAV 614KB 16-bit/44.1kHz,27,28.3,"TTS job queued->finished ~4s; vLLM clean-decode (17+25=42, no BPE artifacts)",3+1,SGD <0.001,"n/a (on-demand mode, not a billable workflow)",2026-06-22,"AI-5b VERIFIED: agent-mode.sh end-to-end PASS — reliable on-demand path confirmed (the failover trigger AI-1 wires to; healthy :8000 = up-signal). From a clean card (2 MiB), agent-mode.sh exited 0 with the AGENT MODE READY banner: vllm-nemotron-agent (NVFP4, 32K ctx, gpu-mem-util 0.82, kv-cache fp8, --max-num-seqs 2, reasoning-parser nemotron_v3, auto tool-choice) + light audio (redis, gateway :9000, RQ worker, Fish Speech :9002). COEXIST HEADLINE: peak 28294 MiB used / 3817 MiB free (~3.8GB headroom) during 3 concurrent vLLM completions + 1 Fish TTS job; steady idle-resident 28162 MiB. Per-proc: vLLM EngineCore 25382 + Fish ~2262 + ~498 MiB. FUNCTIONAL (not just loaded): vLLM clean-decode 17+25->42, finish_reason stop, ZERO BPE artifacts (Mistral-regex tokenizer WARNING benign for this model); Fish TTS job queued->finished ~4s, real 614KB RIFF WAVE 16-bit/44.1kHz mono. /health + /v1/models both 200. restart=no (genuine on-demand), health=healthy. Reasoning-model gotcha re-confirmed (BS-8): 64-tok cap -> empty content/finish_reason length; needs >=256-1024 tok -> content materializes; probes must allow headroom. HARDENING GAP (out of AI-5b scope, flagged): agent-mode.sh stop-list (docker stop vllm-nemotron + grep comfyui|creative-stack|creative-lipsync|creative-music) MISSES vllm-coding (AI-1, same :8000 bind, restart=unless-stopped -> conflict/OOM if running) and the post-06-22 lipsync tiers creative-latentsync/hallo2/musetalk (GPU-capable lazy VRAM). Didn't bite this run (vllm-coding already exited; lipsync stopped pre-flight) so shakedown stands; fix = add those names to the stop-list (EXECUTION-PLAN B2 declarative mode.sh closes the OOM-from-stale-load class). GPU etiquette: owner-freed exclusive window; stopped only what this task started; card returned to free at end.",outputs/audio/voiceover/954c8f67_speech.wav (local only)
CRE-3_ChainVsLoop,FLUX 'maestro' keyframe -> Wan2.2 I2V chain(3 seg) vs loop(1 seg + ping-pong),same scene/seed both duration modes (--duration 10),832x480 24fps MP4 10s each,"chain 1508 (kf 8 + i2v 1500) / loop 494 (kf 4 + i2v 490)",17.9,"chain 3x81fr segments / loop 81fr + ffmpeg ping-pong fill",n,"SGD 0.199 / 0.065","USD 1.50 / 0.60",2026-07-02,"CRE-3 night-batch REPRODUCTION + the missing loop comparison (chain first measured 2026-06-22). CHAIN: 3 full 3.375s segments, drift QA metric-backed this time (cre3_drift_qa): seams are CLEAN (join color_L1 <=1.5, struct MAE 0.0027 both joins - the I2V re-interpretation of the handoff frame is near-invisible) but cumulative drift COMPOUNDS monotonically (last-frame vs seg1-frame0: color_L1 7.8 -> 12.2 -> 14.5; struct MAE 0.028 -> 0.030 -> 0.032): subject scale grows (compounded dolly), palette saturates/warms, silhouette detail erodes by seg 3. No catastrophic artifact this run (vs the 06-22 malformed-gull) - ambient/misty scene is the favorable case. LOOP: zero drift by construction, ping-pong turn + restart frames visually identical, 1/3 the GPU time - but motion REVERSES (walking subject moonwalks), so loop suits static/ambient scenes only. VERDICT unchanged + refined: >3.375s chaining is client-usable for gentle/ambient shots <=3 segments (seams fine, drift is the limit); identity/detail-critical subjects <=2 segments; static scenes -> loop mode at 1/3 cost. GPU during I2V: avg 448.7W / peak 451.8W (450W cap), 72-76C, no thermal throttle, VRAM peak 17.9GB.",samples/cre3-longclip/maestro_chained.mp4
CRE-2_CommercialPipeline_Qwen,Qwen-Image 'sterling' keyframe (Apache-2.0) -> Wan2.2 I2V MoE -> ACE-Step -> mux,scene idea (one line) -> commercial-SAFE full pipeline,832x480 24fps MP4 (3.375s) + score muxed,560,28.1,"560s end-to-end (kf 32s cold + i2v 520s + music 8s)",n,SGD 0.074,USD 0.600,2026-07-02,"CRE-2 residual CLOSED: first full-pipeline run on the commercial-safe DEFAULT (--mode commercial = Qwen-Image keyframe, Apache-2.0 - the 06-22 CRE-2 row used non-commercial FLUX.1-dev and demonstrated only the pipeline shape). Qwen kf 32s (cold-load incl) + Wan2.2 I2V MoE 520s + ACE-Step 8s + mux = 560s; total wall 677s incl the PRISM :8001 CPU expand pre-step. Output commercial_s42_final.mp4 (3.375s, video+audio ffprobe-verified); eye-verified keyframe AND last frame: clean product shot (bottle/orange/mint intact, no melt through I2V). VRAM peak 28.1GB (Qwen kf stage); I2V at 450W avg / 76C peak, no throttle. Every link in this chain is now commercially licensed -> a 'Commercial' full-pipeline tile is warranted.",samples/cre2-commercial/commercial_s42_final.mp4
Scribe-saturated (B-2),Nemotron 3-nano-omni-30b-a3b-reasoning (vLLM 0.20.0 forensic mode),"8 concurrent closed-loop streams, 512-tok outputs, 240s steady window (30s warmup discarded)","sustained saturation throughput + thermal/power envelope",240,29.9,716.8 tok/s aggregate / 89.6 per-stream (sustained 8-stream),8,SGD <0.001,USD ~0.10/1M out,2026-07-02,"B-2 saturation edition (AI-6 night batch): sustained max-load at the B-1 knee concurrency (8) on the FORENSIC config (180K-ctx compose, 29.9GB resident). 336 requests / 172032 tok decoded in the 240s window, ZERO errors. AGG 716.8 tok/s sustained = 98% of the 728 burst knee (B-1 2026-06-13) -> the knee number HOLDS under sustained load, not just burst. Req latency median 5.74s / p90 5.77s (512-tok outputs). POWER avg 405.9W / peak 419W (under the 450W cap); TEMP avg 64C / peak 68C; clocks FLAT ~2865MHz, zero throttle events (SW power cap + HW/SW thermal all Not Active across the window). Complements the 06-22 2-stream agent-config row (437 agg): quote 276 single-stream / 437 @2 / 717 @8 sustained. GPU etiquette: owner-approved serialized window; card handed over clean.",n/a