VERIGEN commited on
Commit
6ee8d39
·
1 Parent(s): 5eb82a5

Krea 2 Raw DiT as INT8 and NF4 .k2q; Qwen-Image VAE fp32 / bf16; Qwen3-VL-4B text encoder bf16 / fp8_scaled

Browse files
README.md CHANGED
@@ -9,38 +9,37 @@ private: true
9
  # VForge-Krea2
10
 
11
  Model files for [vforge](https://github.com/verigen/vforge) (`k2t` trains Krea 2 LoRAs, `k2c` compares them by face
12
- likeness). Private: the files carry upstream licences that apply to every use.
13
 
14
- `k2t model fetch VERIGEN/VForge-Krea2[:ALIAS]` downloads the files listed in `manifest.json`, checks their sha256
15
- and registers the aliases in `~/.config/k2t/config.toml`. The aliases match the defaults `k2t project init` writes,
16
- so a fetched set works with a new project as is.
17
 
18
  ## Layout
19
 
20
  | Path | Alias | Content |
21
  |---|---|---|
22
- | `dit/krea2-raw.int8.k2q.safetensors` | `krea2-int8` | Krea 2 Raw DiT, block Linears pre-quantised to INT8 |
23
- | `dit/krea2-raw.nf4.k2q.safetensors` | `krea2-nf4` | Krea 2 Raw DiT, block Linears pre-quantised to NF4 |
24
- | `vae/qwen-image-vae.safetensors` | `qwen-image-vae` | Qwen-Image VAE |
25
- | `text_encoder/qwen3vl-4b.fp8_scaled.safetensors` | `qwen3vl-4b-fp8` | Qwen3-VL-4B text encoder, fp8_scaled |
26
- | `lora/krea2-turbo-r64.safetensors` | `krea2-turbo-r64` | Krea 2 turbo LoRA, rank 64 (previews and k2c renders) |
27
- | `face/buffalo_l/` | `buffalo_l` | InsightFace detection (`det_10g.onnx`) and recognition (`w600k_r50.onnx`) |
28
- | `compile/<env key>/k2t-cc.zip` | | `k2t compile-cache export` bundle for one GPU model and torch build |
29
-
30
- `.k2q` files load as-is: `k2t model verify FILE` checks the header, layout and checksum. The source sha256, the
31
- quantisation and the format version are in the safetensors header.
32
-
33
- A compile bundle only helps on the exact GPU model and torch / Triton / CUDA build in its env key (`k2t env` prints
34
- the local one). `k2t compile-cache import` refuses a bundle made for another key.
35
-
36
- Versions are this repository's revisions. A changed file replaces the old one in place, and `k2t model fetch
37
- --revision` pins an earlier revision.
38
-
39
- ## Licences
40
-
41
- | Files | Source | Licence |
42
- |---|---|---|
43
- | `face/buffalo_l/` | [InsightFace](https://github.com/deepinsight/insightface) `buffalo_l` | the code is MIT; the pretrained models are for non-commercial research use only |
44
-
45
- The `.k2q` files are converted by vforge from the upstream weights. Each file's row in this table is added together
46
- with the file, and the upstream licence texts go into `LICENSES/`.
 
9
  # VForge-Krea2
10
 
11
  Model files for [vforge](https://github.com/verigen/vforge) (`k2t` trains Krea 2 LoRAs, `k2c` compares them by face
12
+ likeness).
13
 
14
+ `manifest.json` lists every file with its alias and sha256. `k2t model fetch VERIGEN/VForge-Krea2[:ALIAS]` (not
15
+ implemented yet) will download from it, check the hashes and register the aliases in `~/.config/k2t/config.toml`.
 
16
 
17
  ## Layout
18
 
19
  | Path | Alias | Content |
20
  |---|---|---|
21
+ | `dit/krea2-raw.int8.k2q.safetensors` | `krea2-int8` | Krea 2 Raw DiT, block Linears pre-quantised to INT8 (13.5 GB) |
22
+ | `dit/krea2-raw.nf4.k2q.safetensors` | `krea2-nf4` | Krea 2 Raw DiT, block Linears pre-quantised to NF4 (8.2 GB) |
23
+ | `vae/qwen-image-vae.fp32.safetensors` | `qwen-image-vae` | Qwen-Image VAE, fp32: training and the latent cache |
24
+ | `vae/qwen-image-vae.bf16.safetensors` | `qwen-image-vae-bf16` | Qwen-Image VAE, bf16 |
25
+ | `text_encoder/qwen3vl-4b.bf16.safetensors` | `qwen3vl-4b-bf16` | Qwen3-VL-4B text encoder, bf16 (8.9 GB) |
26
+ | `text_encoder/qwen3vl-4b.fp8_scaled.safetensors` | `qwen3vl-4b-fp8` | Qwen3-VL-4B text encoder, fp8_scaled (5.2 GB) |
27
+ | `face/buffalo_l/` | `buffalo_l` | InsightFace detection (`det_10g.onnx`) and recognition (`w600k_r50.onnx`) for `k2c` |
28
+
29
+ `.k2q` files load as-is. `k2t model verify FILE --source SRC` checks the header, layout and data checksum, and
30
+ re-quantises sampled Linears from the source to check that they match bit for bit. Both files here pass it against
31
+ the source below.
32
+
33
+ Compile caches can be added as `compile/<env key>/k2t-cc.zip` (a `k2t compile-cache export` bundle). A bundle only
34
+ hits on the exact GPU model and torch / Triton / CUDA build in its env key, which `k2t env` prints.
35
+
36
+ Versions are this repository's revisions. A changed file replaces the old one in place.
37
+
38
+ ## Sources
39
+
40
+ | Files | Source |
41
+ |---|---|
42
+ | `dit/` | `raw.safetensors` of [krea/Krea-2-Raw](https://huggingface.co/krea/Krea-2-Raw), sha256 `f99bb0ff8e362b77342bc4994e0c50906fe7ef7074864b181b7d48d2fa6d03d7`, converted with `k2t model convert` |
43
+ | `vae/qwen-image-vae.fp32.safetensors`, `text_encoder/qwen3vl-4b.bf16.safetensors` | `vae/` and `text_encoder/` of krea/Krea-2-Raw, unmodified |
44
+ | `vae/qwen-image-vae.bf16.safetensors`, `text_encoder/qwen3vl-4b.fp8_scaled.safetensors` | ComfyUI-layout repacks, unmodified |
45
+ | `face/buffalo_l/` | [InsightFace](https://github.com/deepinsight/insightface) `buffalo_l`; the pretrained models are for non-commercial research use |
dit/krea2-raw.int8.k2q.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:576f988a7f00ebb8967169be5b7e5563dec185ebe6836f8f230fbf0306d76c27
3
+ size 13491960032
dit/krea2-raw.nf4.k2q.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c3a1dafdf49475e61c0b369b743b5ed21f14f32a61076c6b0352da7d52d29686
3
+ size 8167430492
manifest.json CHANGED
@@ -1,6 +1,38 @@
1
  {
2
  "format": 1,
3
  "models": {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4
  "buffalo_l": {
5
  "kind": "face",
6
  "dir": "face/buffalo_l",
 
1
  {
2
  "format": 1,
3
  "models": {
4
+ "krea2-int8": {
5
+ "kind": "dit",
6
+ "quant": "int8",
7
+ "file": "dit/krea2-raw.int8.k2q.safetensors",
8
+ "sha256": "576f988a7f00ebb8967169be5b7e5563dec185ebe6836f8f230fbf0306d76c27"
9
+ },
10
+ "krea2-nf4": {
11
+ "kind": "dit",
12
+ "quant": "nf4",
13
+ "file": "dit/krea2-raw.nf4.k2q.safetensors",
14
+ "sha256": "c3a1dafdf49475e61c0b369b743b5ed21f14f32a61076c6b0352da7d52d29686"
15
+ },
16
+ "qwen-image-vae": {
17
+ "kind": "vae",
18
+ "file": "vae/qwen-image-vae.fp32.safetensors",
19
+ "sha256": "ab1b61103959913d6c7e628cf793dbb2ca4726a40a3b3ae206c52b8e75bf6f08"
20
+ },
21
+ "qwen-image-vae-bf16": {
22
+ "kind": "vae",
23
+ "file": "vae/qwen-image-vae.bf16.safetensors",
24
+ "sha256": "0c8bc8b758c649abef9ea407b95408389a3b2f610d0d10fcb054fe171d0a8344"
25
+ },
26
+ "qwen3vl-4b-bf16": {
27
+ "kind": "text_encoder",
28
+ "file": "text_encoder/qwen3vl-4b.bf16.safetensors",
29
+ "sha256": "8434db05292f95e0041589a7c82abeb39385be59c85b54ae11caa7b45e9f4f13"
30
+ },
31
+ "qwen3vl-4b-fp8": {
32
+ "kind": "text_encoder",
33
+ "file": "text_encoder/qwen3vl-4b.fp8_scaled.safetensors",
34
+ "sha256": "54bd5144df0bbc25dd6ccadfcb826b521445a1b06ae5a42570bdd2974ca87094"
35
+ },
36
  "buffalo_l": {
37
  "kind": "face",
38
  "dir": "face/buffalo_l",
text_encoder/qwen3vl-4b.bf16.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8434db05292f95e0041589a7c82abeb39385be59c85b54ae11caa7b45e9f4f13
3
+ size 8875715136
text_encoder/qwen3vl-4b.fp8_scaled.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:54bd5144df0bbc25dd6ccadfcb826b521445a1b06ae5a42570bdd2974ca87094
3
+ size 5242467968
vae/qwen-image-vae.bf16.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0c8bc8b758c649abef9ea407b95408389a3b2f610d0d10fcb054fe171d0a8344
3
+ size 253806966
vae/qwen-image-vae.fp32.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ab1b61103959913d6c7e628cf793dbb2ca4726a40a3b3ae206c52b8e75bf6f08
3
+ size 507591892