downgrade model version

4 days ago · 94a69a4053
6 changed files with 364 additions and 29 deletions
--- a/API-GUIDE.md
+++ b/API-GUIDE.md
@ -18,7 +18,7 @@
 ## Обзор
-DaSiWa API — асинхронный REST API для генерации видео из изображений с использованием DaSiWa WAN 2.2 Lightspeed моделей через ComfyUI.
+DaSiWa API — асинхронный REST API для генерации видео из изображений с использованием DaSiWa WAN 2.2 TastySin v8.1 моделей через ComfyUI.
 **Архитектура:** Submit → Poll → Retrieve (как RunPod)
@ -247,7 +247,7 @@ Health check сервера. **Не требует аутентификации.
 - **Размеры:** должны быть кратны 16
 - **Length:** рекомендуется кратно 8 + 1 (например: 81, 89, 97)
- **Steps:** DaSiWa Lightspeed оптимизирован под 4 шага (можно больше, но медленнее)
+- **Steps:** DaSiWa TastySin оптимизирован под 4 шага (можно больше, но медленнее)
 - **CFG:** значения > 2.0 могут давать артефакты
 ---
@ -475,12 +475,6 @@ if queue_size > 5:
 - Недостаточно VRAM
 - Workflow файл повреждён
 ### Ошибка "Promotion for Float8 Types is not supported"
 - **Причина:** Lightspeed модели несовместимы с fp8 text encoder
 - **Решение:** Используй fp16 text encoder (`umt5_xxl_fp16.safetensors`)
 - Официальные WAN 2.2 модели работают с fp8, но кастомные Lightspeed файнтюны требуют fp16
 ### Медленная генерация
 - Проверьте загрузку GPU: `nvidia-smi`
@ -496,7 +490,7 @@ if queue_size > 5:
 - ✨ Endpoints: `/run`, `/status`, `/purge`
 - ✨ Background worker thread
 - ✨ Queue management
- 🔧 Обновлён на DaSiWa WAN 2.2 Lightspeed
+- 🔧 Обновлён на DaSiWa WAN 2.2 TastySin v8.1
 - 🔧 Упрощён workflow (14 нод вместо 50+)
 ### v1.0 (2026-03-06)
--- a/README.md
+++ b/README.md
@ -1,6 +1,7 @@
 # 🎬 DaSiWa API Server для ComfyUI
-Автономный API сервер для генерации видео через ComfyUI на выделенном GPU сервере (Intelion Cloud и т.д.).
+Автономный API сервер для генерации видео через ComfyUI на выделенном GPU сервере.
 Модель: **DaSiWa WAN 2.2 TastySin v8.1** (I2V / FLF2V, 4 шага, Lightx2v дистилляция).
 **Безопасность:** HMAC подпись каждого запроса (timestamp + nonce + body). Перехват бесполезен — подпись уникальна.
@ -15,10 +16,9 @@ custom_comfyui/
 ├── hmac_auth.py        # HMAC авторизация (нужен и там, и там)
 ├── generate_keys.py    # Генерация ключей (один раз)
 ├── setup.sh            # Автоустановка: Python, зависимости, UFW, systemd
 ├── dasiwa-api.service  # Systemd сервис (автозапуск)
 ├── requirements.txt    # Python зависимости
 ├── keys.json           # 🔒 Ключи (НЕ коммитить!)
-└── workflow_api.json   # 🎨 ComfyUI workflow (DaSiWa WAN 2.2 Lightspeed)
+└── workflow_api.json   # 🎨 ComfyUI workflow (TastySin v8.1)
 ```
 ---
@ -28,36 +28,36 @@ custom_comfyui/
 ### 1. Загрузка моделей на сервер
 ```powershell
-$SERVER = "user@<ip_сервера>"
+$SERVER = "root@194.67.95.42"
-# HIGH checkpoint (~13 GB)
+# HIGH checkpoint (~9.7 GB)
-scp "DasiwaWAN22I2V14BLightspeed_synthseductionHighV9.safetensors" ${SERVER}:/ComfyUI/models/checkpoints/
+scp "TastySin-HIGH-v8.1.safetensors" ${SERVER}:/ComfyUI/models/checkpoints/
-# LOW checkpoint (~13 GB)
+# LOW checkpoint (~9.7 GB)
-scp "DasiwaWAN22I2V14BLightspeed_synthseductionLowV9.safetensors" ${SERVER}:/ComfyUI/models/checkpoints/
+scp "TastySin-LOW-v8.1.safetensors" ${SERVER}:/ComfyUI/models/checkpoints/
 # VAE (~335 MB)
 scp "wan_2.1_vae.safetensors" ${SERVER}:/ComfyUI/models/vae/
-# Text Encoder (~9.4 GB, fp16 - required for Lightspeed models)
+# Text Encoder (~4.9 GB, fp8)
-scp "umt5_xxl_fp16.safetensors" ${SERVER}:/ComfyUI/models/text_encoders/
+scp "umt5_xxl_fp8_e4m3fn_scaled.safetensors" ${SERVER}:/ComfyUI/models/text_encoders/
 ```
 **Скачать модели:**
- **Checkpoints:** [DaSiWa WAN 2.2 i2v 14B (S) Lightspeed](https://civitai.com/models/1981116)
+- **Checkpoints:** [DaSiWa WAN 2.2 TastySin v8.1](https://civitai.com/models/1329763) — скачай HIGH и LOW версии (Safetensors)
 - **VAE:** [wan_2.1_vae.safetensors](https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/blob/main/split_files/vae/wan_2.1_vae.safetensors)
- **Text Encoder:** [umt5_xxl_fp16.safetensors](https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/resolve/main/split_files/text_encoders/umt5_xxl_fp16.safetensors) (fp16 required for Lightspeed models)
+- **Text Encoder:** [umt5_xxl_fp8_e4m3fn_scaled.safetensors](https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/resolve/main/split_files/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors)
 **Структура моделей на сервере:**
 ```
 📂 ComfyUI/models/
 ├── checkpoints/
-│   ├── DasiwaWAN22I2V14BLightspeed_synthseductionHighV9.safetensors
+│   ├── TastySin-HIGH-v8.1.safetensors
-│   └── DasiwaWAN22I2V14BLightspeed_synthseductionLowV9.safetensors
+│   └── TastySin-LOW-v8.1.safetensors
 ├── vae/
 │   └── wan_2.1_vae.safetensors
 └── text_encoders/
-    └── umt5_xxl_fp16.safetensors
+    └── umt5_xxl_fp8_e4m3fn_scaled.safetensors
 ```
 ### 2. Генерация ключей (на любом ПК с Python)
--- a/server.py
+++ b/server.py
@ -1,6 +1,6 @@
 #!/usr/bin/env python3
 """
-DaSiWa I2V/FLF2V API Server для ComfyUI.
+DaSiWa I2V/FLF2V API Server для ComfyUI (TastySin v8.1).
 Работает рядом с ComfyUI на той же машине.
 Асинхронный API (как RunPod):
--- a/setup.sh
+++ b/setup.sh
@ -177,5 +177,5 @@ echo "🧪 Проверка:"
 echo "   curl http://localhost:${API_PORT}/health"
 echo ""
 echo "🔐 API доступен на:"
-echo "   http://<ip_сервера>:${API_PORT}/generate"
+echo "   http://<ip_сервера>:${API_PORT}/run"
 echo "============================================================"
--- a/synthSeductionRelease.html
+++ b/synthSeductionRelease.html
@ -0,0 +1,341 @@
 <div>
 <div>
 <article>
 <div>
 <div>
 <h1>Greetings USER #333777 - Registered Master!</h1>
 <h2>DaSiWa WAN 2.2 "SynthSeduction" arrives!</h2>
 <p><em>New breakthroughs in adopting virtual high efficient and neuronal entities transformed into human-like appearance were made.</em></p>
 <p><em>Please register, Master. The new 1st class Synth will be ready to serve you!</em></p>
 <hr/>
 <p>This time the whole package will be delivered to you Safetensors | GGUF (Q8 - Q2) and the TrueVision format in a wholesome Synth, ready to serve your will!</p>
 <video><source/><source/></video>
 <p>I'm continuing in enhancing my DaSiWa WAN 2.2 i2v checkpoint series with this one as a direct successor of v8.</p>
 <p>I made some really awesome progression in the past and not every time the jump can be equally big, but let's focus on the new 💎💎💎 shiny augmentations! 💎💎💎<br/>This time the release focus on aspects that stand on my wish-list for a long while plus some quality improvements and&nbsp;a needed (from my view)&nbsp;tuning down of some bias's to a more general approach.</p>
 <p>This also means that some or another really specialised NSFW content may need a bit more prompting or a lora to be perfect, but that also means, that the understanding/motions will not be pushed as much into some unwanted situations!</p>
 <p>Will this make things worse, you ask? - No! 💮 It will perform superior in many situations and will even be better with spacial effects in general that were not possible, while just be slightly worse on some minor situations! Less bias is sometimes a good thing to maintain overall quality. 🧬 Fine-tuning is always a process!</p>
 <p>👁️&zwj;🗨️ The major focus this time:</p>
 <ul>
 <li>
 <p>❤️&zwj;🔥Raised prompt adherence</p>
 </li>
 <li>
 <p>💫 Faster and better motions, without sacrificing quality (~ 20-30% more motion)</p>
 </li>
 <li>
 <p>🪄 Best in-class details even on low resolutions</p>
 </li>
 <li>
 <p>💎 Consistent details over the full video (almost never detail shift)</p>
 </li>
 <li>
 <p>🎊 Better special effects from prompts</p>
 </li>
 <li>
 <p>Lower bias</p>
 </li>
 <li>
 <p>Reduced slow-motions</p>
 </li>
 <li>
 <p>Raise lora compatibility</p>
 </li>
 <li>
 <p>Maintain overall NSFW understanding</p>
 </li>
 </ul>
 <hr/>
 <p>💠 With this release I also build on my foundation to expand the DaSiWa - Series with a new approach! <br/>💠 "TrueVision" !</p>
 <p>This aims to be the quality you know from my checkpoints without any techniques that reduces quality in any way. Like distillation loras, iteration speed-up's or merges that will reduce quality by any means I can test.</p>
 <p>This will have the most pure version of my checkpoint. The goal is to enable you to mix and match technologies as you like (FFGO, Infinity, custom distillations,...) with the look and feel of my other optimisations.</p>
 <p>Due to the nature of the long generation times with this approach, TrueVision will not be as extensively showcased/tested by me, but will give you more freedom to use it.</p>
 <hr/>
 <h2>Here are some technical facts about the new checkpoint:</h2>
 <ul>
 <li>
 <p>Only open-source resources</p>
 </li>
 <li>
 <p>Fast, with only 4 steps - lightx2v r64-1022 distillation (latest/best)</p>
 </li>
 <li>
 <p>As always tested with many situations (NSFW/SFW/realistic/anime)</p>
 </li>
 <li>
 <p>Full FP16 base</p>
 </li>
 <li>
 <p>Baked with CFGZeroStar</p>
 </li>
 <li>
 <p>Baked SkyReels V2 for longer video consistency</p>
 </li>
 <li>
 <p>Baked with reward attention for best guessing of the situation</p>
 </li>
 <li>
 <p>Baked VTA (Better completions, transitions, input understanding)</p>
 </li>
 <li>
 <p>Pushed scaling on fine-grained tuning parameters to the limits for optimised results</p>
 </li>
 <li>
 <p>Highly compatible with LoRAs</p>
 </li>
 <li>
 <p>Excluded CLIP</p>
 <ul>
 <li>
 <p>Better control over offloading and less memory dependent from the view of the single checkpoint if used with custom clip</p>
 </li>
 <li>
 <p>smaller size obviously</p>
 </li>
 </ul>
 </li>
 </ul>
 <ul>
 <li>
 <p>Zero prompt i2v are possible! The understanding is really awesome! Try it! 😉</p>
 </li>
 </ul>
 <h3>What to expect from this release:</h3>
 <ul>
 <li>
 <p>Superior details even with low resolution - including anime/realistic images (lips, eyes, ears, breasts, genitals, ...)!</p>
 </li>
 <li>
 <p>Higher/Faster natural and smooth motion</p>
 </li>
 <li>
 <p>High success on guesstimating details</p>
 </li>
 <li>
 <p>Overall better results</p>
 </li>
 <li>
 <p>Better compatibility with LoRAs</p>
 </li>
 <li>
 <p>Usable with your preferred/custom CLIP (if compatible)</p>
 </li>
 <li>
 <p>Again reduced hallucinations</p>
 </li>
 <li>
 <p>Best in-class prompt adherence</p>
 </li>
 <li>
 <p>Better texts</p>
 </li>
 <li>
 <p>Better camera control</p>
 </li>
 <li>
 <p>Less bias</p>
 </li>
 <li>
 <p>Stable NSFW understanding</p>
 </li>
 </ul>
 <p>Because examples show more than 1000 words:</p>
 <p>See for yourself - the release will showcase various examples on the checkpoint page ~</p>
 <ul>
 <li>
 <p>in various situations</p>
 </li>
 <li>
 <p>SFW and NSFW</p>
 </li>
 <li>
 <p>Anime, semi-realistic, realistic (illu)</p>
 </li>
 <li>
 <p>All samples without any LoRA</p>
 </li>
 <li>
 <p>Initial images (anime) made with my Illustrious checkpoint</p>
 </li>
 <li>
 <p>Made with my workflow</p>
 </li>
 <li>
 <p>All same seed 333777 (Master-Registration-Code 😁)</p>
 </li>
 <li>
 <p>All showcases are 1st try, to showcase the capabilities! (no cherry picking)</p>
 </li>
 <li>
 <p>All include metadata/workfow</p>
 </li>
 <li>
 <p>Settings:</p>
 <ul>
 <li>
 <p>Same prompts, seeds, initial images</p>
 </li>
 <li>
 <p>Metadata included</p>
 </li>
 <li>
 <p>Complex scenes</p>
 </li>
 <li>
 <p>anime style, realistic style</p>
 </li>
 </ul>
 </li>
 <li>
 <p>A brief comparison see below 👇</p>
 <ul>
 <li>
 <p>SAFETENSOR (FP8+)</p>
 </li>
 <li>
 <p>GGUF (Q8-Q2)</p>
 </li>
 <li>
 <p>Basic WAN 2.2 with just lightx2v lora</p>
 </li>
 </ul>
 </li>
 </ul>
 <p>As always this is meant to be as versatile as possible, but enhance the understanding of WAN 2.2 in NSFW and not trained situations, maintaining best in-class quality.</p>
 <p><em>Want to make NSFW? Realistic? Semi-Realistic? Anime? ... all in! 💯</em></p>
 <p>The goal remains to create something useful and simple, yet powerful in understanding and consistency.</p>
 <h3>What not to expect:</h3>
 <ul>
 <li>
 <p>The checkpoint is not meant to replace every LoRA-Concept out there, it will be great on its own and even more awesome with a specialised LoRA!<br/>So make sure to support the awesome LoRA-Creators!</p>
 </li>
 <li>
 <p>Arbitrary perfect guessing of NSFW details out of visibility in every situation (WAN 2.2 is not trained on that at the end)</p>
 </li>
 <li>
 <p>To be made for concepts that are against the ToS like "nudifying" real individuals. If this works out by coincidence, it would simply be a basic understanding of WAN 2.2.</p>
 </li>
 <li>
 <p>Guessing the lower genitals in every situation and style, you may need a LoRA. (Forcing this into the model results just in a total bias to the chosen gender)</p>
 </li>
 <li>
 <p>To understand every situation and NSFW concept possible</p>
 </li>
 </ul>
 <hr/>
 <h3>Comparison</h3>
 <p>All the same prompt, same SEED and resolution.</p>
 <video><source/><source/></video> <video><source/><source/></video><hr/>
 <h3>Workflow</h3>
 <p>There is a fitting optimised workflow on my latest conclusions, so make sure to try it out! 😸</p>
 <p>Remember. WAN 2.2 is not perfect, neither any merge/fine-tune will be. So there are always some bad samples.</p>
 <hr/>
 <h3>Support and early access</h3>
 <p>I hope you have fun and you can bring your ideas to life with that awesome tool! 😉</p>
 <p>As I'm doing this in my spare time and I'm limited on resources, consider buying me a ko-fi, if you want to support me. Maybe Buzz / EA, everything helps!</p>
 <p>All checkpoints will be available through EA and free after.</p>
 <p>In the end it will be your decision to spend some buzz and support further development or wait till the EA ends.</p>
 <hr/>
 <h3>Closing words and sources</h3>
 <p>💌 Spacial thanks to the resources, videos, and works of all these awesome creators!</p>
 <p>Make sure to check out their stuff! ✨</p>
 <p>Listed resources may be partly used or part of the process to create that checkpoint if open source or free of use:</p>
 <ul>
 <li>
 <p>https://huggingface.co/Comfy-Org</p>
 </li>
 <li>
 <p>https://huggingface.co/lightx2v</p>
 </li>
 <li>
 <p>https://huggingface.co/alibaba-pai</p>
 </li>
 <li>
 <p>https://huggingface.co/RaphaelLiu</p>
 </li>
 <li>
 <p>https://huggingface.co/Kijai/WanVideo_comfy/tree/main</p>
 </li>
 <li>
 <p>https://github.com/SkyworkAI</p>
 </li>
 <li>
 <p>https://civitai.com/user/darksidewalker</p>
 </li>
 <li>
 <p>https://civitai.com/user/TwoMoreLurker</p>
 </li>
 <li>
 <p>https://civitai.com/user/CubeyAI</p>
 </li>
 <li>
 <p>https://civitai.com/user/imb101</p>
 </li>
 <li>
 <p>https://civitai.com/user/alcaitiff</p>
 </li>
 <li>
 <p>https://civitai.com/user/dngstn32</p>
 </li>
 <li>
 <p>https://civitai.com/user/takoyakioji</p>
 </li>
 <li>
 <p>https://github.com/kijai/ComfyUI-KJNodes</p>
 </li>
 <li>
 <p>https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite</p>
 </li>
 <li>
 <p>https://github.com/Artificial-Sweetener/comfyui-WhiteRabbit</p>
 </li>
 <li>
 <p>https://github.com/Fannovel16/ComfyUI-Frame-Interpolation</p>
 </li>
 <li>
 <p>https://github.com/rgthree/rgthree-comfy</p>
 </li>
 <li>
 <p>https://github.com/SeanScripts/ComfyUI-Unload-Model</p>
 </li>
 <li>
 <p>https://github.com/ggml-org/llama.cpp</p>
 </li>
 <li>
 <p>https://github.com/city96/ComfyUI-GGUF</p>
 </li>
 <li>
 <p>https://github.com/silveroxides/convert_to_quant</p>
 </li>
 <li>
 <p>Own python scripts</p>
 </li>
 <li>
 <p>Own resources</p>
 </li>
 </ul>
 </div>
 </div>
 </article>
 <div>
 <div>
 <p>👍</p>
 <button>01234567890123456789</button>
 <p>❤️</p>
 <button>0123456789</button>
 <p>😂</p>
 <button>0123456789</button>
 <p>😢</p>
 <button>0123456789</button></div>
 <div>
 <div>
 <div>
 <p>10</p>
 </div>
 </div>
 <div>
 <p>10</p>
 </div>
 <button></button></div>
 </div>
 </div>
 </div>
--- a/workflow_api.json
+++ b/workflow_api.json
@ -1,7 +1,7 @@
 {
  "1": {
    "inputs": {
-      "ckpt_name": "DasiwaWAN22I2V14BLightspeed_synthseductionHighV9.safetensors"
+      "ckpt_name": "TastySin-HIGH-v8.1.safetensors"
    },
    "class_type": "CheckpointLoaderSimple",
    "_meta": {
@ -10,7 +10,7 @@
  },
  "2": {
    "inputs": {
-      "ckpt_name": "DasiwaWAN22I2V14BLightspeed_synthseductionLowV9.safetensors"
+      "ckpt_name": "TastySin-LOW-v8.1.safetensors"
    },
    "class_type": "CheckpointLoaderSimple",
    "_meta": {
@ -28,7 +28,7 @@
  },
  "4": {
    "inputs": {
-      "clip_name": "umt5_xxl_fp16.safetensors",
+      "clip_name": "umt5_xxl_fp8_e4m3fn_scaled.safetensors",
      "type": "wan",
      "device": "default"
    },