-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathproviders.yaml
More file actions
145 lines (135 loc) · 4.84 KB
/
Copy pathproviders.yaml
File metadata and controls
145 lines (135 loc) · 4.84 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
# Model and compute registry, researched 2026-07-20.
# Swapping a model or host = editing this file, never workflow code.
# Prices in AUD, converted from provider USD rates at AUD/USD 0.70
# (2026-07-20). Rates drift; re-verify before a big batch.
# Licence notes matter because the music universe is commercial.
runners:
local:
kind: local
command: null # set when a local worker script exists (Phase 1)
remote_pod:
kind: remote_pod
host: null # fill in when a pod is up, e.g. root@1.2.3.4
port: 22
remote_dir: /workspace/jobs
command: "cd {remote_job_dir} && python /workspace/worker.py --spec spec.json --out results"
fal:
kind: saas
service: fal.ai
# Where to rent the metal (2026-07-20 snapshot):
# draft batches: Vast.ai interruptible RTX 4090 ~A$0.45/hr (per-second
# billing, docker-native, built for restartable batch)
# reliability up: RunPod community 4090 ~A$0.50/hr, secure A100 80GB
# ~A$2.00/hr, H100 ~A$4.30/hr; network volumes keep
# model weights warm between runs; Sydney region exists
# zero-ops: fal.ai per-output (Wan hosted ~A$0.07/s of video)
# for premium clips without managing weights
models:
video:
draft:
name: HunyuanVideo 1.5
licence: Apache 2.0
vram_gb: 14-24
why: best quality per GB on a rented 4090/5090; ComfyUI native
standard:
name: Wan 2.2 A14B
licence: Apache 2.0
vram_gb: 24 (FP8) / 40-80 (FP16)
why: best open quality with a clean commercial licence; huge LoRA ecosystem
premium:
name: LTX-2.3
licence: LTX-2 Community (free commercial under US$10M/yr revenue,
their licence term is set in USD)
vram_gb: 16-80
why: native synced audio+video, native 4K and vertical 1080x1920
lipsync:
draft:
name: MuseTalk 1.5
licence: MIT-style
vram_gb: 6-8
why: realtime-class, cheap bulk dubbing passes
standard:
name: LatentSync 1.x
licence: Apache 2.0
vram_gb: 8-12
why: best pure lip fidelity for dubbing existing footage
premium:
name: InfiniteTalk
licence: Apache 2.0
vram_gb: 24
why: unlimited-length audio-driven performance, head and body sync
avatar:
premium:
name: HunyuanVideo-Avatar
licence: Tencent community (check terms before commercial use)
vram_gb: 24-80
why: one image + audio to full-body emotion-controllable presenter
tts:
draft:
name: Kokoro-82M
licence: Apache 2.0
vram_gb: cpu-ok
why: fast fixed-voice narration; cannot clone
standard:
name: Qwen3-TTS
licence: Apache 2.0
vram_gb: 4-8
why: 3-second-reference cloning, commercial-safe default
premium:
name: Chatterbox Multilingual v3
licence: MIT
vram_gb: 6-8
why: beat ElevenLabs in blind tests; note output watermarking
avoid:
- "F5-TTS (CC-BY-NC), Fish Speech (CC-BY-NC-SA), XTTS-v2 (CPML):
all non-commercial, excluded from this pipeline"
image:
draft:
name: FLUX.2 Klein 4B
licence: Apache 2.0
vram_gb: 8-10
why: the only permissive FLUX.2 weight; commercial-safe drafts
standard:
name: Qwen-Image + Qwen-Image-Edit 2511
licence: Apache 2.0
vram_gb: 12-24
why: 2026 standard for character consistency via edit-with-reference;
best open text rendering for comic lettering and signage
premium:
name: FLUX.2 dev
licence: NON-COMMERCIAL open weights; paid licence for commercial
vram_gb: 32 (FP8) - 64
why: quality leader, 10 native reference images for character lock;
only usable here with a paid BFL licence
character_lock: >
Recurring cast: train a character LoRA on Qwen-Image or FLUX.1-dev,
roughly 30-60 min on one rented A100/4090 (about A$2-7 per character).
llm:
draft:
name: Qwen3.6-27B (Q4)
licence: Apache 2.0
vram_gb: 16-17
why: strongest single-24GB-card ideation model
premium:
name: GPT-OSS-120B
licence: Apache 2.0
vram_gb: 64-80
why: A100/H100-tier scripting brain for batch ideation on a pod
hosted:
name: claude CLI (already subscribed)
why: zero extra cost for interactive ideation; used by brief --run-claude
embeddings:
standard:
name: Qwen3-Embedding-0.6B
licence: Apache 2.0
why: cheap local semantic search over the vault
premium:
name: Qwen3-Embedding-8B
licence: Apache 2.0
why: tops open MTEB v2 when retrieval quality matters
beat_sync:
tools:
- librosa + ffmpeg (beat/onset extraction driving cut lists)
- Demucs (stem separation, key visuals off drums or vocals)
- ComfyUI_Yvann-Nodes (audio-reactive generation weights)
- lum3on/ComfyUI_AudioTools (BPM + frame-synced beat events)