feat: add MiniT2I support (#1683)

2026-07-23 11:20:53 -05:00 · 2026-07-02 00:46:51 +08:00
parent 1a13107bac
commit 3590aa8d62
13 changed files with 1006 additions and 12 deletions
--- a/docs/minit2i.md
+++ b/docs/minit2i.md
@@ -0,0 +1,48 @@
+# How to Use
+
+MiniT2I uses a MiniT2I diffusion transformer and `google/flan-t5-large` as the text encoder.
+
+## Download weights
+
+- Download MiniT2I diffusion model
+    - safetensors: https://huggingface.co/MiniT2I/MiniT2I/tree/main/minit2i-b-16/transformer (`diffusion_pytorch_model.safetensors`)
+- Download flan-t5-large text encoder
+    - safetensors: https://huggingface.co/google/flan-t5-large/tree/main (`model.safetensors`)
+
+## Examples
+
+### Mac Metal
+
+```
+./bin/sd-cli \
+  --backend metal \
+  --diffusion-model ../models/minit2i/diffusion_pytorch_model.safetensors \
+  --t5xxl ../models/flan-t5-large/model.safetensors \
+  --prompt "a cat" \
+  --steps 100 \
+  --cfg-scale 6 \
+  --width 512 \
+  --height 512 \
+  --seed 42 \
+  --sampling-method euler \
+  --rng cpu \
+  --output minit2i_metal.png \
+  --threads 8
+```
+
+### CUDA with diffusion flash attention
+
+```
+./bin/sd-cli \
+  --diffusion-model ../models/minit2i/diffusion_pytorch_model.safetensors \
+  --t5xxl ../models/flan-t5-large/model.safetensors \
+  --prompt "a cat" \
+  --steps 100 \
+  --cfg-scale 6 \
+  --width 512 \
+  --height 512 \
+  --seed 42 \
+  --sampling-method euler \
+  --diffusion-fa \
+  --output minit2i_cuda.png
+```