Promote the condition_proj-weighted matrices to the root, move the unweighted ones to obsolete/
Browse files
README.md
CHANGED
|
@@ -45,18 +45,30 @@ It works because every Qwen3-VL shares the **same tokenizer** (151936 tokens): a
|
|
| 45 |
|
| 46 |
Put them in `ComfyUI/models/clip_projections/`.
|
| 47 |
|
| 48 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 49 |
|---|---|---|---|
|
| 50 |
-
| `
|
| 51 |
-
| `
|
| 52 |
-
| `
|
| 53 |
-
| `
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
\* **measured in a different space β not comparable to the rows below.** See the next section.
|
| 60 |
|
| 61 |
The **8B matrices need an 8B encoder** β different input width, 4096 instead of 2560, so they are not interchangeable with the 4B ones. The node checks the width and refuses a mismatch.
|
| 62 |
|
|
|
|
| 45 |
|
| 46 |
Put them in `ComfyUI/models/clip_projections/`.
|
| 47 |
|
| 48 |
+
**Take the four files at the root.** Everything in `obsolete/` is superseded and kept only so the comparison can be reproduced.
|
| 49 |
+
|
| 50 |
+
| File | Use it for | Encoder | Corpus | Test cosine |
|
| 51 |
+
|---|---|---|---|---|
|
| 52 |
+
| `h3_qwen3vl_4b_CONDPROJ_tap24.pt` | **text prompts, the default** | any Qwen3-VL-4B | 240 prompts | 0.845 * |
|
| 53 |
+
| `h3_qwen3vl_8b_CONDPROJ_tap24.pt` | **ref2va and image references** | any Qwen3-VL-8B | 203 prompts | 0.860 * |
|
| 54 |
+
| `h3_control_identity.pt` | control β run it once | β | β | β |
|
| 55 |
+
| `h3_control_zero.pt` | control β run it once | β | β | β |
|
| 56 |
+
|
| 57 |
+
\* measured in a different space from the rows below β **the two sets of numbers cannot be compared**. See the next section.
|
| 58 |
+
|
| 59 |
+
<details>
|
| 60 |
+
<summary><code>obsolete/</code> β the unweighted matrices, kept for comparison</summary>
|
| 61 |
+
|
| 62 |
+
| File | Encoder | Corpus | Test cosine |
|
| 63 |
|---|---|---|---|
|
| 64 |
+
| `obsolete/h3_qwen3vl_4b_tap24.pt` | `qwen3vl_4b_bf16` | 2 000 prompts | 0.712 |
|
| 65 |
+
| `obsolete/h3_qwen3vl_4b_tap24_200prompts.pt` | `qwen3vl_4b_bf16` | 200 prompts | 0.699 |
|
| 66 |
+
| `obsolete/h3_qwen3vl_4b_int8convrot_tap24.pt` | `qwen3vl_4b_int8_convrot` | 240 prompts | 0.697 |
|
| 67 |
+
| `obsolete/h3_qwen3vl_8b_tap24.pt` | `qwen3vl_8b_nvfp4` | 203 prompts | 0.731 |
|
| 68 |
+
|
| 69 |
+
These were the v0.1.1 release. They work, and the 2 000-prompt one is calibrated on eight times more data than anything at the root β but they minimise the wrong error, and the root matrices are the ones that generation testing actually ran on. Keep them if you want to A/B the weighting yourself.
|
| 70 |
+
|
| 71 |
+
</details>
|
|
|
|
|
|
|
| 72 |
|
| 73 |
The **8B matrices need an 8B encoder** β different input width, 4096 instead of 2560, so they are not interchangeable with the 4B ones. The node checks the width and refuses a mismatch.
|
| 74 |
|
obsolete/h3_qwen3vl_4b_int8convrot_tap24.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:11aa0b0f14c8dd5a2f0c2206d3a44dabebbe4167ffb13bee0c23817d2ceab587
|
| 3 |
+
size 52514229
|
obsolete/h3_qwen3vl_4b_tap24.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:941b94675714985232346ffb9e06cf8a3dfb1c61cfa6e2fb24a733dabec60df1
|
| 3 |
+
size 52514021
|
obsolete/h3_qwen3vl_4b_tap24_200prompts.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:74a78d637e3799bd380c772c016e7a8e577216699c191d703c331d9c541cda8b
|
| 3 |
+
size 52514089
|
obsolete/h3_qwen3vl_8b_tap24.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:363d15189c5a5fdd67d831e613108e9e9babd88fd2d09e6a14727493f89b4f33
|
| 3 |
+
size 83983589
|