SD3.5-Large-GGUF-mixed-sdcpp

Overview

These models are made to work with stable-diffusion.cpp release master-ac54e00 onwards. Support for other inference backends is not guarenteed.

Quantized using this PR https://github.com/leejet/stable-diffusion.cpp/pull/447

Normal K-quants are not working properly with SD3.5-Large models because around 90% of the weights are in tensors whose shape doesn't match the 256 superblock size of K-quants and therefore can't be quantized this way. Mixing quantization types allows us to take adventage of the better fidelity of k-quants to some extent while keeping the model file size relatively small.

Only the second layers of both MLPs in each MMDiT block of SD3.5 Large models have the correct shape to be compatible with k-quants. That still makes up for about 10% of all the parameters.

Files:

Non-Linear Type:

• sd3.5_large-iq4_nl.gguf: Same size as q4_k_4_0 and q4_0, runs faster than q4_k_4_0 (on Vulkan at least), and provides image quality somewhat comparable to q5_1 model. Recommended

Mixed Types:

• sd3.5_large-q2_k_4_0.gguf: Smallest quantization yet. Use this if you can't afford anything bigger
• sd3.5_large-q3_k_4_0.gguf: Degraded, but usable at high step count.
• sd3.5_large-q4_k_4_0.gguf: Exacty same size as q4_0 and iq4_nl, I recommend using iq4_nl instead.
• sd3.5_large-q4_k_4_1.gguf: Smaller than q4_1, and with comparable degradation. Recommended
• sd3.5_large-q4_k_5_0.gguf: Smaller than q5_0, and with comparable degradation. Very close to the original f16 already. Recommended

Legacy types:

• sd3.5_large-q4_0.gguf: Same size as q4_k_4_0, Not recommended (use iqk_nl or q4_k_4_0 instead)
• sd3.5_large-q4_1.gguf: Not recommended (q4_k_4_1 is better and smaller)
• sd3.5_large-q5_0.gguf: Barely better and bigger than q4_k_5_0
• sd3.5_large-q5_1.gguf: Better and bigger than q5_0
• sd3.5_large-q8_0.gguf: Basically indistinguishable from the original f16, but much smaller. Recommended for best quality

Outputs:

Sorted by model size (Note that q4_0 and q4_k_4_0 are the exact same size)

only 28 steps, cfg scale 4.5

Generated with a modified version of sdcpp with this PR applied to enable clip timestep embeddings support.

Text encoders used: q4_k quant of t5xxl, full precision clip_g, and q8 quant of ViT-L-14-TEXT-detail-improved-hiT-GmP-TE-only-HF in place of clip_l.

Full prompts and settings in png metadata.