Hardware

What your rig can actually do.

Every gaming GPU can make AI videos — but the results vary. Here's an honest breakdown of typical performance across common setups. Times are approximate; run your own benchmarks before committing to a workflow.

Seeded from the Hybrig reference fleet (James's actual rigs). Crowdsourced benchmarks will extend this.

flagship

Bigglesworth-Studios

The flagship rig — runs everything locally, trains your LoRAs overnight.

≈ $3,200 new

GPU
NVIDIA RTX 4090 (24GB)
CPU
AMD Ryzen 7 7800X3D (8C / 4.2 GHz)
RAM
32 GB DDR5 @ 4800 MT/s
OS
Windows 11

Best for

  • +Wan 2.7 I2V video generation at 720p–1080p
  • +Personal LoRA training (4–8 hrs, ~1GB file)
  • +FLUX.2 dev + Z-Image Turbo image gen at full quality
  • +Qwen-Image-Edit 2511 logo swaps + scene edits in seconds

Not good for

  • −LTX-2.3 at 4K — possible but slow; 1080p is the sweet spot at 24GB

Typical render times

Wan 2.7 I2V, 8s, 720p~3–5 min
Wan 2.7 I2V, 8s, 1080p~6–9 min— edge of VRAM; use fp8 quant if OOM
FLUX.2 dev image, 1024×1024~10–15 sec
Z-Image Turbo, 1024×1024<1 sec
Qwen-Image-Edit 2511 logo swap, 1MP~5–10 sec
Personal LoRA training (Wan)~4–8 hrs— overnight run
Whisper-small, 60s audio~2–3 sec

Role in a Hybrig fleet

Primary renderer. Handles heavy Wan + LoRA. If you only have one machine, this is the one.

Worker role local

Local models: wan-2.7, flux-2-dev, z-image-turbo, qwen-image-edit-2511, whisper

workstation

Bigglesworth

The workhorse — handles image gen and lighter video, offloads from the flagship.

≈ $1,400 new

GPU
NVIDIA RTX 3080 (12GB)
CPU
AMD Ryzen 7 3800X (8C / 3.9 GHz)
RAM
32 GB DDR4 @ 2666 MT/s
OS
Windows 11 Pro

Best for

  • +Z-Image Turbo + Qwen-Image 2.0 image gen
  • +Wardrobe-lock + inpainting batch runs
  • +Wan 2.7 video with fp8/GGUF quantization (720p, a bit slower)
  • +HyperFrames headless Chrome rendering

Not good for

  • −Wan 2.7 at full precision — OOM likely at 1080p, quantization required
  • −FLUX.2 dev — too big at 12GB; Z-Image Turbo covers the everyday need
  • −LoRA training runs — technically possible at low-rank but impractical (20+ hrs)

Typical render times

Wan 2.7 I2V, 8s, 720p (quantized)~5–8 min— fp8/GGUF quant required
Wan 2.7 I2V, 8s, 1080pnot supported— OOM — use flagship or cloud
Z-Image Turbo, 1024×1024~1–2 sec
Qwen-Image-Edit 2511 (GGUF Q4)~15–30 sec— tight at 12GB — Q4 + Lightning LoRA
Personal LoRA training~20+ hrs— impractical at this tier
Whisper-small, 60s audio~3–5 sec

Role in a Hybrig fleet

Image gen + wardrobe batch. Runs in parallel with flagship for 2x throughput on mixed workloads.

Worker role local

Local models: z-image-turbo, qwen-image-2, wan-2.7 (quantized)

utility

The Utility Box (rename me)

The utility box — transcription, face embedding, post-production CPU-bound work.

≈ $600 new

GPU
NVIDIA RTX 2070 (8GB)
CPU
older desktop CPU
RAM
16 GB typical
OS
Windows / Linux

Best for

  • +Whisper transcription (all sizes up to medium)
  • +Face embedding / ArcFace lookups
  • +HyperFrames headless-Chrome rendering
  • +Image upscaling (Clarity / Real-ESRGAN)

Not good for

  • −Any video generation — VRAM too low
  • −FLUX.2 / Qwen at full quality — use a quantized Z-Image Turbo

Typical render times

Wan 2.7 video generationnot supported— VRAM below threshold
Z-Image Turbo (quantized), 1024×1024~5–10 sec— quantized variant
Face embedding (ArcFace)<1 sec
Whisper-medium, 60s audio~5–8 sec
HyperFrames render, 30s clip~2–3 min
Image upscale 2x (Real-ESRGAN)~10–15 sec

Role in a Hybrig fleet

Support box. Runs the parts flagship shouldn't waste cycles on — transcription, embedding, post.

Worker role local

Local models: z-image-turbo (quantized), whisper, face-embed-local (planned)

mobile

Mobile Station (MacBook Pro M4 Pro)

The mobile station — full local image gen + editing on unified memory; video rides the cloud.

≈ $2,800 new

GPU
Apple M4 Pro 20-core GPU (unified memory)
CPU
Apple M4 Pro 14-core CPU
RAM
24–48 GB unified memory
OS
macOS Sequoia

Best for

  • +Z-Image Turbo + Qwen-Image 2.0 image gen on MPS (GGUF, no CUDA needed)
  • +Qwen-Image-Edit 2511 logo swaps — the Logo Swap tool runs fully local here
  • +Whisper transcription on Apple Neural Engine
  • +UI authoring — the primary client-side creative console when traveling

Not good for

  • −Wan 2.7 / LTX-2.3 video generation at 24GB unified — needs 32GB+ or an NVIDIA card
  • −FP8 checkpoints — Metal can't load them; always use GGUF quants
  • −LoRA training for video models — requires CUDA ecosystem

Typical render times

Wan 2.7 I2V via Metal (24GB)not supported— GGUF path exists but ~40+ min per 2s clip — use cloud
Z-Image Turbo, 1024×1024~15–30 sec— MPS-optimized runners available
Qwen-Image-Edit 2511 logo swap (GGUF Q4 + Lightning)~1–3 min— close heavy apps during batches
Whisper-large, 60s audio~2–4 sec— Neural Engine accelerated
Personal LoRA trainingnot supported— CUDA-only workflow
HyperFrames render, 30s clip~1–2 min

Role in a Hybrig fleet

Mobile authoring + local image lab. Image gen and logo swaps run on-device; video jobs route to the fleet or cloud. 32-48GB configs unlock local video too.

Worker role local

Local models: z-image-turbo, qwen-image-2, qwen-image-edit-2511, whisper

cloud-only

No GPU / cloud-only

No GPU? Hybrig still works — cloud-only path, pay per render.

no GPU required

GPU
None (or integrated graphics)
CPU
any modern CPU
RAM
8+ GB
OS
Windows / macOS / Linux

Best for

  • +Anyone without a dedicated GPU
  • +Travelers / phone users opening Hybrig on the move
  • +Users who want zero local setup

Not good for

  • −The cost savings story. Every draft hits the cloud — pays ~$1.94/take for Seedance Fast, $3-5 for finals.
  • −Offline work

Typical render times

Seedance Fast, 8s, 720p~3–5 min— $1.94 per take
Seedance Pro, 8s, 1080p~4–8 min— $3–5 per take
Flux via fal, 1024×1024~10–20 sec— ~$0.04 per image
ElevenLabs voice cloningcloud-native— $22/mo subscription
Personal LoRA trainingnot supported— requires local GPU
HyperFrames render, 30s clipcloud TBD— requires server-side headless Chrome

Role in a Hybrig fleet

The zero-setup path. Everything runs on fal / ElevenLabs / Gemini. No local worker needed.

Worker role cloud

Don't see your hardware? Benchmarks are approximate — your setup probably falls between two of these profiles. Once you're running, Hybrig's Dashboard shows actual wall-clock times on your own jobs for real calibration.

Setup local GPU rendering →Back to Dashboard