← All of today's comparisons

Edits: Qwen 2.1 base vs H3 chroma vs H3 Pruna

4 edit cases at seed 42 on the V100, in the first picture's aspect: Qwen at its ~1 MP, H3 at 1 and 2 MP, and Pruna also at 4 MP the way its edit workflow does it (picture 1 scaled to 4 MP with lanczos first; Pruna's README says edits work best at 4 or 8 MP). The 2 MP cards come from the earlier run (same prompts and seed; times are whole warm jobs, not split). Each model set starts unloaded (ComfyUI /free, model files not in the OS cache), so its first job (outfit) includes loading everything from disk; that job then runs again warm. "Loading" = cold minus warm. "Diffusion" = the sampling steps alone, from ComfyUI's per-step progress events. RAM→VRAM is what still moves between system RAM and the GPU during a warm job (H3's 21 GB DiT and its text encoder don't both fit in 32 GB of VRAM). Click a picture to open the viewer: ← / → flips between versions, Z or click toggles 1:1 pixels.

cold jobsame job warmloadingdiffusiontext encoderVAE decodeRAM→VRAM warm
Qwen 2.1 base · turbo 8 steps · 1 MP54.2 s19.4 s34.8 s11.7 s (8 × 1.46)4.9 s0.5 s0 GB
H3 chroma · 1 MP169.0 s57.7 s111.3 s19.4 s (6 × 3.24)6.9 s0.8 s24 GB
H3 Pruna · 1 MP135.8 s35.0 s100.9 s21.2 s (8 × 2.65)4.5 s8.0 s9 GB
H3 Pruna · 4 MP167.6 s92.8 s74.8 s74.4 s (8 × 9.30)3.8 s8.9 s32 GB

Warm job time per case (all models already loaded):

OutfitSign textAnime restyleTwo pictures (cat into shop)
Qwen 2.1 base · turbo 8 steps · 1 MP19 s (diffusion 12)19 s (diffusion 12)19 s (diffusion 12)28 s (diffusion 14)
H3 chroma · 1 MP58 s (diffusion 19)59 s (diffusion 20)55 s (diffusion 19)75 s (diffusion 31)
H3 chroma · 2 MP73 s72 s72 s89 s
H3 Pruna · 1 MP35 s (diffusion 21)36 s (diffusion 21)36 s (diffusion 21)52 s (diffusion 36)
H3 Pruna · 2 MP41 s41 s41 s59 s
H3 Pruna · 4 MP93 s (diffusion 74)93 s (diffusion 74)92 s (diffusion 74)116 s (diffusion 95)
Qwen 2.1 base · turbo 8 steps · 1 MP (bot /edit default) H3 chroma · 1 MP (bot model h3chroma) H3 chroma · 2 MP (earlier run) H3 Pruna · 1 MP (Pruna LoRA + encoder, SOL VAE) H3 Pruna · 2 MP (earlier run) H3 Pruna · 4 MP (Pruna's edit size) source picture