What changed
Every data refresh and site update, newest first.
2026-10-11Image editing measured on the RTX 5060 Ti 16 GB: FLUX.1 Kontext [dev] FP8 (53 s per edit) and GGUF Q8_0 (82 s), and Qwen-Image-Edit 2511 GGUF Q4_K_M with the 4-step Lightning LoRA (27 s). Free workflows for all three, a before-and-after gallery, and a new guide comparing the two editors. Model pages now link their tested workflow.2026-10-10Out-of-memory guide extended: why VRAM stays full after a run, real memory leaks, and slow runs caused by shared GPU memory. New guides: the best local image model for 16 GB (9 models, same prompt, timed, with licences) and which of the 50 models you can use commercially (open, with conditions, non-commercial), with the smallest card each fits on.2026-10-10New: a page for each of the 74 shared ComfyUI files (text encoders, VAEs, vision encoders) — what it is, which folder it goes in, its size, other versions, and which models use it. File names in every download list now link to these pages.2026-10-06New: 12 free ComfyUI workflows, each one run end to end on my RTX 5060 Ti 16 GB — FLUX.1 dev and schnell, FLUX.2 klein 4B and 9B, Z-Image Turbo, Qwen-Image 2.1, Krea 2, Ming-Image and Wan 2.2 5B and 14B — with a side-by-side gallery of what each makes from the same prompt. Each workflow opens with a note listing every file, its folder and the measured speed. Model pages link the workflow next to the download list. Social preview images per model.2026-10-04Measured on my RTX 5060 Ti 16 GB with ComfyUI 0.38.1: Qwen-Image 2.1 (INT8 17.2 s, 16-bit 39.7 s, GGUF Q8_0 43.6 s), Krea 2 Turbo (INT8 10.9 s, FP8 17.7 s) and Ming-Image (INT8 7.4 s, 16-bit 17.4 s) per 1024×1024 image. INT8 was the fastest 8-bit file, so on RTX 40/50 and RX 9000 the calculator now tries INT8 right after FP8 and before Q8_0 GGUF.2026-10-01Measured on my RTX 5060 Ti 16 GB: FLUX.2 klein 4B (FP8 2.9 s, 16-bit 3.9 s, GGUF Q8_0 4.3 s per 1024×1024 image) and klein 9B (GGUF Q8_0 8.5 s, Q4_K_M 9.8 s). New guides: everything that runs on an RTX 5060 Ti 16 GB, 12 GB vs 16 GB of VRAM, and FLUX.2 klein timed.2026-10-01Monthly data check: all 875 file sizes in 110 Hugging Face repos re-read, no changes. New model: Ming-Image 0.1 Design (inclusionAI, 6B, MIT) — small image model, very large text encoder. New GPU: Radeon 8065S (Ryzen AI Max+ PRO 495, Gorgon Halo) with up to 160 GB of unified memory for the GPU.2026-09-27Pages where a model does not fit well now show one clearly marked affiliate link to rent a bigger GPU on Vast.ai. It never affects a verdict. Details on the About page.2026-09-27Measured on my RTX 5060 Ti 16 GB with 32 GB RAM: Wan 2.2 14B text to video (832×480, 81 frames) as FP8, GGUF Q5_K_M and GGUF Q8_0. FP8 does not fit in 16 GB but was the fastest. New guide: Wan 2.2 14B on a 16 GB card.2026-09-27First measured results from my own RTX 5060 Ti 16 GB: FLUX.1 [dev] as FP8, GGUF Q8_0 and GGUF Q4_K_S, FLUX.1 [schnell] Q8_0, Z-Image Turbo BF16 (11 s per 1024×1024 image) and Wan 2.2 5B FP16 (about 8.5 minutes for a 121-frame 1280×704 clip), with time and peak VRAM. New guide: FP8 vs GGUF on a 16 GB RTX 5060 Ti.2026-09-25Apple Silicon Macs: 12 memory sizes from 16 to 512 GB (M1 to M5 and the new M6), with the GPU share of memory measured on current macOS, which file types work on Apple GPUs (FP8 does not), GGUF text encoders for Macs, real reported speeds and a Mac guide page. The checker recognises Macs too.2026-09-25New tools: a VRAM calculator (resolution, frames, batch, LoRAs, ControlNets, system RAM), "what would an upgrade change" for any two GPUs plus 35 ready-made upgrade pages, and a LoRA training page with 49 figures from the trainers' own docs. Every answer page now lists exactly what to download — model, text encoder, VAE and extra files, with direct links, total size and the system RAM you need. Model pages show which file types are fast on NVIDIA, AMD and Intel. Wan pages estimate blocks_to_swap for Kijai's WanVideoWrapper. The checker can suggest your GPU from your browser and remember it (stored only in your browser). New models page with email alerts and a feed. Fixed: the red card-limit line on memory maps was drawn in the wrong place.2026-09-259 more models (SCAIL-2, Boogu-Image, ERNIE-Image, HiDream-O1-Image, Mage-Flow, Lumina Image 2.0, HiDream-I1 Full, the original HunyuanVideo 13B, LTX-Video 13B) — 49 in total. 69 reported real-world results with sources, including AMD, Intel Arc, laptops and Strix Halo.2026-09-25Fact-check pass: FP8 compute on AMD RX 9000 (RDNA 4) added; ComfyUI Dynamic VRAM (default on NVIDIA since March 2026) now explained on verdict pages; --lowvram advice updated; FP4 speed figure corrected to NVIDIA's actual numbers; LTX-2.5 text encoder corrected to Gemma 4 12B.2026-09-25New design. 14 more models (Krea 2, Qwen-Image 2.1, MiniMax H3, Ideogram 4, Z-Image, Wan 2.1 I2V/VACE, Wan 2.2 Animate/S2V, Wan Animate 2, FLUX.1 Fill, SD 1.5, Illustrious/Pony). 33 laptop GPUs and Strix Halo. GPU comparisons, "best GPU for" pages, memory maps on every verdict, site search. All 530 original file sizes re-checked against Hugging Face: zero mismatches.2026-09-24First version: 25 models, 53 desktop GPUs, 886 pages.