5s I2V to 30fps
Generate ~5 seconds at 10 fps (50 frames), then interpolate to 30 fps (150 frames). Mapped against Hermies + bigscreen ComfyUI (RX 9070, ~16GB).
You do not need a model that outputs 150 frames natively. Generate a coherent 50-frame clip, then use frame interpolation (RIFE / FILM) to triple the frame rate. The hard part is the I2V pass — not encoding or Immich upload.
| Step | What runs | Fits today? |
|---|---|---|
| 1. Source | Immich still via immich-get / Hermies |
Yes — already wired |
| 2. Image to video | Wan-class I2V, ~50 frames @ 10fps, 480p–720p | No — need longer model + more memory than SVD/9070 |
| 3. Interpolate | 10fps → 30fps (insert 2 frames between each pair) | Yes on 16GB+ after I2V unloads |
| 4. Deliver | mp4 encode → Immich COMFY.GEN | Yes — same pattern as comfy-i2v |
Bottleneck is step 2. Estimates assume ~50 frames at ~480p–720p with a Wan-class 14B I2V (FP8/quant where noted). Unified-memory mini PCs have huge capacity but lower bandwidth than discrete GPUs.
| Platform | Memory | 5s @ 10fps I2V | Interp → 30fps | Fit for this job |
|---|---|---|---|---|
| RX 9070 (current bigscreen) | 16GB VRAM | Too tight / unreliable | Fine | Keep for SVD shorts; not Wan 50-frame |
| RTX 4090 / 5090 | 24GB VRAM | Doable at ~480p FP8 | Fine | Practical minimum discrete upgrade |
| RTX A6000 / 6000 Ada | 48GB VRAM | Comfortable, 720p headroom | Fine | Best discrete workstation pick |
| DGX Spark (GB10) | 128GB unified | Capacity yes; slower than 24–48GB discrete | Fine | One box for LLM + I2V; CUDA-native |
| AMD Ryzen AI Max+ 395 (Strix Halo) | 128GB unified | Capacity yes; ROCm/Comfy video less mature | Fine | Spark-class alternative; cheaper, x86; more Wan friction |
| H100 / dual 24GB | 80GB / 48GB | Easy | Fine | Overkill for 5s; longer/higher-res later |
Both are compact 128GB unified-memory deskside systems. Spark wins on CUDA / Comfy maturity for video diffusion. Strix Halo (e.g. Framework Desktop, HP Z2 Mini G1a class) is usually cheaper and runs normal x86 Windows/Linux, but Wan I2V on ROCm is more DIY than CUDA. Neither replaces a 4090 for raw diffusion speed — they win on model size fit and sharing memory with an LLM on the same machine.
| Layer | Stack | vs Hermies today |
|---|---|---|
| Orchestration | Wrapper like comfy-i2v (Immich in/out) |
Pattern exists; new command/workflow |
| I2V | Wan 2.1/2.2 I2V 14B FP8 + Comfy API graph | Not installed (SVD only, ~14–25 frames) |
| Interpolation | RIFE Comfy node or standalone | Not wired; light memory |
| Encode / upload | H.264 mp4 → immich-put → COMFY.GEN |
Already proven |
SVD is trained for ~14–25 frames. Stretching to 50 does not give clean 5s of motion — quality falls apart or you OOM. Interpolation only smooths what you already generated; it cannot invent missing seconds. Wan-class I2V is the length engine; RIFE is the fps polish.
Discrete path: 24GB+ NVIDIA on the Comfy machine, Wan I2V + RIFE, keep Hermies/Ollama on glifftop.
Unified path: DGX Spark (easiest CUDA) or Strix Halo 128GB (budget / x86), run LLM and video on one box with careful memory scheduling.
Occasional use: cloud I2V may beat buying hardware. Local wins for Immich/Hermies privacy and frequent LAN runs. Current smoke-tested path remains SVD ~2s @ 6fps on the 9070.
Local open-source estimates, not vendor quotes · 2026-08-25 · Spark/Strix Halo rows are capacity-oriented; measure throughput on a real Wan workflow before buying for video alone.