Update model card
Browse files
README.md
CHANGED
|
@@ -19,13 +19,41 @@ Every LoRA ships as `<Name>_KREA2.safetensors` alongside `<Name>_KREA2.md` insid
|
|
| 19 |
|
| 20 |
### Repository restructure
|
| 21 |
- All characters moved under a single **`Characters/`** directory (one folder per character containing the `.safetensors`, the persona `.md`, and the training captions).
|
|
|
|
| 22 |
- A new **`Pruned_Comparison`** folder was added with one preview image per character (see below).
|
| 23 |
|
| 24 |
### Dynamic re-rank + block pruning
|
| 25 |
-
Every LoRA was re-ranked with **dynamic `sv_fro` SVD** (retention `0.998`, max rank `16`) and block-pruned by dropping the **8 lowest-impact transformer blocks** (blocks `1-8`, ~3% of delta energy).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
|
| 27 |
### Retrained character
|
| 28 |
-
- **Jung Ho-yeon** (`Ho-yeon`) was **retrained** on a new dataset, then re-ranked and pruned like the others.
|
|
|
|
|
|
|
| 29 |
|
| 30 |
## Fully automated pipeline
|
| 31 |
|
|
@@ -40,39 +68,6 @@ This repository - dataset, training, and upload - was produced end-to-end by an
|
|
| 40 |
7. **Training** - LoRAs trained on pods with **Ostris AI-Toolkit** with user predefined settings.
|
| 41 |
8. **Publishing** - weights and persona files uploaded to this repo by the agent.
|
| 42 |
|
| 43 |
-
## Contents
|
| 44 |
-
|
| 45 |
-
| | | |
|
| 46 |
-
|---|---|---|
|
| 47 |
-
| Adria Arjona | Adriana Lima | Alexandria Ocasio-Cortez |
|
| 48 |
-
| Alix Lapri | Amelia Gray | Ana De Armas |
|
| 49 |
-
| Anne Hathaway | Anya Taylor-Joy | Ariana Grande |
|
| 50 |
-
| Autumn Falls | Barbara Palvin | Bella Poarch |
|
| 51 |
-
| Billie Eilish | Camila Cabello | Camille Cerf |
|
| 52 |
-
| Camille Hurel | Cara Delevingne | Carla Poquin |
|
| 53 |
-
| Cindy Kimberly | Corinna Kopf | Daniela Avanzini |
|
| 54 |
-
| Demitra Kalogeras | Doja Cat | Dorozea |
|
| 55 |
-
| Dua Lipa | Eiza González | Elydia |
|
| 56 |
-
| Emilia Clarke | Emily Ratajkowski | Eva Mendes |
|
| 57 |
-
| Fernanda Ly | Gal Gadot | Hailey Rose |
|
| 58 |
-
| Hannah Kleit | Ho-yeon | Ice Spice |
|
| 59 |
-
| Imaan Hammam | Inde Navarrette | Irina Shayk |
|
| 60 |
-
| Jessica Alba | Kaia Gerber | Kelsey Merritt |
|
| 61 |
-
| Kendall Jenner | Kylie Jenner | Lara Rajagopalan |
|
| 62 |
-
| Lauren Mayberry | Lea Elui | Leah Halton |
|
| 63 |
-
| Lena Gercke | Lena Mahfouf | Léa Meniel |
|
| 64 |
-
| Madalina Ghenea | Madison Beer | Malu Trevejo |
|
| 65 |
-
| Manon Bannerman | Margot Robbie | Megan Fox |
|
| 66 |
-
| Megan Skiendiel | Monica Bellucci | Natalie Portman |
|
| 67 |
-
| Nicki Minaj | Nikita Bellucci | Olivia Cooke |
|
| 68 |
-
| Pokimane | Rihanna | Rima Hassan |
|
| 69 |
-
| Rosalia | Sab Zada | Scarlett Johansson |
|
| 70 |
-
| Shakira | Shay | Sofia Jamora |
|
| 71 |
-
| Sofia Richie Grainge | Sommer Ray | Sophia Laforteza |
|
| 72 |
-
| Sydney Sweeney | Tinashe | Tokyo |
|
| 73 |
-
| Tyla | Valkyrae | Wejdene |
|
| 74 |
-
| Yoonchae Jeong | Zendaya | Zoe Kravitz |
|
| 75 |
-
|
| 76 |
## Pruned comparison
|
| 77 |
|
| 78 |
One example image per character, generated with the pruned LoRA. Click a character to expand.
|
|
|
|
| 19 |
|
| 20 |
### Repository restructure
|
| 21 |
- All characters moved under a single **`Characters/`** directory (one folder per character containing the `.safetensors`, the persona `.md`, and the training captions).
|
| 22 |
+
- All **`Images_Prompt_Examples`** folders and their contents were **removed**.
|
| 23 |
- A new **`Pruned_Comparison`** folder was added with one preview image per character (see below).
|
| 24 |
|
| 25 |
### Dynamic re-rank + block pruning
|
| 26 |
+
Every LoRA was re-ranked with **dynamic `sv_fro` SVD** (retention `0.998`, max rank `16`) and block-pruned by dropping the **8 lowest-impact transformer blocks** (blocks `1-8`, ~3% of delta energy). All `txtfusion` refiner/layerwise keys are preserved. Weights are stored as `lora_up` / `lora_down` + `alpha` in fp16, loadable directly in ComfyUI. Each file carries `crt_*` provenance metadata.
|
| 27 |
+
|
| 28 |
+
### Trigger word corrections
|
| 29 |
+
27 persona `.md` files had trigger words that did not match the tokens actually used in their training captions. These were corrected to the exact training token, e.g.:
|
| 30 |
+
|
| 31 |
+
| Character | Was | Now |
|
| 32 |
+
|---|---|---|
|
| 33 |
+
| Adriana Lima | `Adriana Lima` | `AdrianaLima` |
|
| 34 |
+
| Anne Hathaway | `Anne Hathaway` | `AnneHathaway` |
|
| 35 |
+
| Ariana Grande | `Ariana Grande` | `ArianaGrande` |
|
| 36 |
+
| Billie Eilish | `Billie_ilish` | `BillieEilish` |
|
| 37 |
+
| Doja Cat | `Dojcat` | `DojaCat` |
|
| 38 |
+
| Elydia | `unknown` | `Helydia` |
|
| 39 |
+
| Eva Mendes | `unknown` | `EvaMendes` |
|
| 40 |
+
| Ho-yeon | `ho-yeon` | `JungHoYeon` |
|
| 41 |
+
| Jessica Alba | `Jessica_lba` | `JessicaAlba` |
|
| 42 |
+
| Kylie Jenner | `Kylie_enner` | `KylieJenner` |
|
| 43 |
+
| Lea Elui | `Léa-elui` | `LeaElui` |
|
| 44 |
+
| Nicki Minaj | `NickiM` | `NickiMinaj` |
|
| 45 |
+
| Sab Zada | `SZA` | `SabZada` |
|
| 46 |
+
| Shakira | `Sh_akira` | `Shakira` |
|
| 47 |
+
| Shay | `Vanessa Lesnicki` | `Shay` |
|
| 48 |
+
| Tokyo | `TKY` | `Tokyo` |
|
| 49 |
+
| Zoe Kravitz | `Zkravit` | `ZoeKravitz` |
|
| 50 |
+
|
| 51 |
+
(plus Alexandria Ocasio-Cortez, Ana De Armas, Anya Taylor-Joy, Bella Poarch, Camille Cerf, Dua Lipa, Emilia Clarke, Hannah Kleit, Inde Navarrette, Kaia Gerber)
|
| 52 |
|
| 53 |
### Retrained character
|
| 54 |
+
- **Jung Ho-yeon** (`Ho-yeon`) was **retrained** on a new dataset, then re-ranked and pruned like the others. Its trigger word is now **`JungHoYeon`**.
|
| 55 |
+
|
| 56 |
+
> Note: `Autumn Falls` and `Hailey Rose` have no matching training dataset, so their trigger word remains `unknown`.
|
| 57 |
|
| 58 |
## Fully automated pipeline
|
| 59 |
|
|
|
|
| 68 |
7. **Training** - LoRAs trained on pods with **Ostris AI-Toolkit** with user predefined settings.
|
| 69 |
8. **Publishing** - weights and persona files uploaded to this repo by the agent.
|
| 70 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 71 |
## Pruned comparison
|
| 72 |
|
| 73 |
One example image per character, generated with the pruned LoRA. Click a character to expand.
|