929721518本人嘅qq細群號，有咩唔識關於tpony嘅問題可以入嚟問。記得備註c站吖

模型已經內置vae，唔需要額外添加vae

The model already has included vae, there is no need to add additional vae

最佳出圖策略係適中分辨率開高清修復，而唔係直接用大分辨率直出

The best generate strategy is to use high-fix at a moderate resolution, rather than directly using high-resolution direct output

[未認證]Tonade正喺創作T-ponynai3模型作者，c站id：Tonade， | 愛發電 (afdian.net)

呢度係愛發電嘅贊助通道，覺得模型好用又有餘力嘅話可以支持下！千祈唔好勉強，多謝你哋每一份支持，會繼續探索點樣練靚模型！

(33) T-ponynai3-v5 - （權重修改版本） | Stable Diffusion Checkpoint | 吐司 tusi.cn (tusiart.com) tusiart(中國版 tensor) 線上生成連結

(由於模型只能同時喺Tusi同Tensor存在，建議係Tusi使用。如有使用問題，請多多指出）

v5版本新增咗4個style，可以透過style_1到style_4微調畫面細節（理論上係咁，但實際效果較玄學）

V5 version 新增咗4個style，可以用style_1至style_4細調圖像細節（理論上係咁，但實際效果較神秘或較弱）

本模型完美支持以ponyv6為底模訓練嘅模型，ani3，sdxl1.0嘅Lora亦可喺某程度上適配

This model perfectly supports lora trained with ponyv6 as the base model, and the Lora of ani3 and sdxl1.0 can also be adapted to some extent.

基於v4.1嘅圖生圖測試（之前版本忽略咗呢部分）

Image inpaint testing based on v4.1 (this is a previously overlooked part)

pony係神，兼容性滿分。模型支持ani，pony嘅lora

必備前置效果詞同ponydiffusion一樣

positive:(score_9,score_8_up,score_7_up,score_6_up,score_5_up,score_4_up)

或者 (score_9,score_8_up,score_7_up)

負面亦可加：

negative: (score_4,score_3,score_2,score_1),

亦可加一般nai系負面詞，例如：

negative: worst quality, bad hands, bad feet

hope u like it ᕕ(◠ڼ◠)ᕗ 基於nai3同ponyv6

訓練須知：v1用咗94張圖片，v2用119張，v3用348張，v3.5用474張，nai3生成嘅圖，用嚟訓練lora融入底模微調，pony支持嘅畫師tag全部都支持，超過兩個畫師tag可能會導致背景崩潰，現時已發現可以生成原神角色，其他未清楚，對呢模型嘅測試唔多，驚嘆於佢對nai3畫風嘅復刻。底模係T-anime-xl、ponyv6同ani3融合模型，未曾公開。

訓練時用嘅係我自己嘅3090顯卡，由v1到v3分別用咗7小時、12小時、35小時同47小時

訓練說明：融合了用94張圖訓練嘅v1、119張圖嘅v2、348張圖嘅v3、474張圖嘅v3.5，全部由NAI3生成，用嚟對底模做微調。Pony支持所有ponyv6已有嘅畫師tag，但nai3無新增任何畫師tag。用超過兩個畫師tag可能導致背景崩潰。現時覺得可以生成原神角色，其他角色未測試。對模型嘅測試唔多，驚歎於佢對NAI3畫風嘅還原度。底模係T-anime-xl同ponyv6同ani3嘅融合模型，尚未發佈。

我用嘅訓練顯卡係自己嘅3090，v1到v3.5分別用咗7小時、12小時、35小時同47小時。

v1

一次有趣嘅嘗試

An interesting attempt

v2

喺v1嘅基礎上略微增加咗訓練集，經過約30小時嘅參數試錯，但繪畫風格仍有部分過擬合，例如雙肚臍眼同雜亂嘅頭髮

On the basis of v1, the training set was slightly increased and went through about 30 hours of trial and error, but the trained art style still had some overfitting, such as double navel eyes and messy hair

v3

v3嘅肢體表现比v2更好，對footfocus嘅理解更深，可以生成視覺衝擊大嘅腳，並呈現更難嘅透視。v3嘅頭髮嘅AI感比v2弱，因v2訓練集太少，所以頭髮稍有過擬合，v2曾出現嘅雙肚臍眼亦消失。整體而言，三倍於v2嘅訓練集同更大dim參數令畫風更自然，長prompt表現遠優於v2。

The limbs of v3 are better than those of v2. In terms of understanding footfocus, v3 can generate feet with greater visual impact and higher difficulty perspective. The AI feeling of v3's hair is also weaker than that of v2, because v2 has too little training set, so the hair part may be slightly overfitting, and the occasional double navel eyes that appear in v2 are also gone. Overall, three times the size of the v2 training set and a larger dim parameter make the art style fit more natural, and the performance is much stronger than v2 under long prompts.

v3.5

喺呢個版本，對質素詞嘅要求唔係咁嚴格，可以完全唔使用pony嘅美學評分質素詞去生成圖片。測試時偶爾會出現無意義色塊，只需將美學評分嘅質素詞換成1.5版常用質素詞，例如score_1, score_2換成worst quality。我喺此版本多加咗約150張訓練集用嚟平衡同充實畫風，並減低學習曲線嘅初始斜率，令模型冇咁容易過擬合，能適配更多lora同奇思妙想嘅提示詞。總體上，呢版本係比v3更自由嘅版本，對男性刻畫遠勝v3，有部分提示詞下嘅色彩同畫風唔會咁鮮艷同油膩。

In this version, the requirements for quality words are not so strict, you can completely not to use the quality words of pony's aesthetic score to plot the picture, and occasionally there will be a situation where the picture generates meaningless color blocks in the test, you only need to replace the quality words of the aesthetic score with 1.5 commonly used quality words, such as score_1, score_2 replace it with worst quality. In this version, I added about 150 more training sets to balance and enrich the art style, and reduced the initial slope of the learning curve, which makes this model less overfitted and can be adapted to more lora and whimsical prompts. Overall, this version is a freer version than the v3 version, and this version is much stronger than the v3 version, and the colors and style of painting under some hints are not so bright and greasy.

v4

呢個版本用798張圖作為訓練素材，用3090顯卡訓練咗90小時。相比v3.5，喺部分prompt下嘅構圖同部分身體部位刻畫更準確，例如手指重影同身體某些部位重疊。prompt方面，我主要以中長度同稍短嘅prompt做訓練目標，始終無人鍾意要寫長長串prompt先有好圖。去除pony嘅美學得分質素prompt後，圖像質素明顯提升，畫面偏平面而非立體，更接近經典動漫風格。ponyv6微調對圖像數量嘅測試亦接近尾聲，接落嚟會由prompt訓練標籤入手，嘗試喺pony有限單次訓練素材量中加入更多可調prompt（例如加入美學評分，目前訓練仍用主流質素詞覆蓋掉pony嘅美學質素詞），並持續增添合適新素材，例如場景素材同更多足部素材（v4嘅足部素材似乎較少）。

This version used 798 images as training materials and trained for 90 hours using a 3090 graphics card. This version has a more accurate composition and depiction of certain parts in certain prompts compared to v3.5, such as ghosting of fingers and overlapping of some body parts. In terms of prompts, my main training goal is to use medium and slightly shorter prompts, as nobody likes to write a long string of prompts to generate high-quality images, right? After removing the quality prompt of Pony's aesthetic score, the image quality has been significantly improved compared to v3.5, and the resulting quality tends to be more flat rather than three-dimensional, closer to the classic anime style. The testing of the fine-tuning effect of Ponyv6 on the number of images is nearing completion. The next step is to start with the training labels of prompts and try to add more adjustable prompts to Pony's limited number of single training materials (such as adding aesthetic scores, the current training logic still uses mainstream quality words to cover Pony's aesthetic score quality words), and continue to add suitable new training materials, such as scene training materials and more foot training materials (v4's foot training materials seem to be a bit scarce).

v4.1

向各位用戶道歉，咁短時間內又出新版本，呢點極考驗電腦內存同網絡速度。O_O

Firstly, I would like to apologize to all users for the release of a new version in such a short period of time, which greatly tests the computer's memory and network speed. O_O

呢版本係基於v4嘅肢體調試版本，因為v4嘅肢體效果較難控制，手部完美率測試期間未達預期。所以我同朋友木貓貓貓對v4作出調整改善，最終令v4.1嘅肢體達到預期，我會釋放幾張xy圖清晰顯示v4.1同v4喺相同參數下嘅生成圖改進。

This new version is based on the limb debugging version of v4. Due to the difficulty in controlling the limb effects of v4, the perfection rate of the hands did not meet my testing expectations in the past few days. So my friend 木猫猫猫 and I made some adjustments and improvements to v4, which ultimately made the limbs of v4.1 meet my expectations. I will release several xy graphs to clearly show the improvement of v4.1 compared to v4 under the same parameters.

v5

呢版本訓練素材減少咗，因為v4失利所以我展開另一項目，從低顯存佔用嘅角度測試想法，訓練咗四個唔同畫風適配T-ponynai3嘅lora，當然原始模型都已上傳至civitai。適配性測試完成後，我開始將呢四組唔同畫風作為添加劑訓練入T-ponynai3-v5。令人驚喜係v5嘅線條質感大大提升，可能因為我訓練咗非常細膩嘅素材。呢四個畫風嘅標記我用咗style_1至style_4提示詞，可惜唔知點解，呢四個風格未能各自分開，或者效果微弱，反而好好融合咗原始風格。雖然未達成多畫風支持目標，但成功提升咗原始nai3畫風嘅質感，或下版本嘗試更進一步。（我好鍾意打機，訓練時唔俾玩電腦遊戲對我嚟講好難）

The training materials for this version have been reduced. Due to the failure of v4, I launched another project to test my idea from a small perspective of memory usage, which is to train four different art styles of Lora adapted to T-ponynai3. Of course, the original model was also uploaded to Civitai. After testing the adaptability, I started training these four different art styles as additives into T-ponynai3-v5. Surprisingly, The line texture of v5 has improved to a high level, probably because I trained a very delicate material. For the marking of these four art styles, I used the prompt words from style_1 to style_4. Unfortunately, for some reason, these four art styles were not separated or the effect was weak, but rather integrated well into the original art style. Although it did not achieve the goal of supporting multiple art styles, it effectively elevated the texture of the original Nai3 art style to a higher level. Perhaps the next version can try to take it even further. (I really enjoy playing games, and it's too difficult for me to play computer games every time I train.)

就v5版本嘅部分問題作總結。

1，lora兼容性同肢體及模糊眼睛問題。lora兼容性因為呢次訓練我用嘅最終權重較高，部分情況會過擬合。呢個優化版本係減低咗相應權重嘅版本，肢體崩壞率同部分lora兼容性有望改善，我跑咗幾張基於v4.1訓練嘅畫風lora對比圖作參考。模糊眼問題應係因為我訓練咗style_1，原始素材眼睛本身模糊，透過style_3或4可改善。

2，體積光曝光問題。我測試時冇見過呢問題，應係我用咗noise offset參數，令模型對光相關提示詞嘅敏感度提升，令原本相同權重嘅光提示詞出來嘅結果更明亮。我建議試下唔用括號及數字提升權重，因sdxl對提示詞敏感，反覆同一提示詞多次不易導致極端結果。同時用呢參數為咗修正提示詞少時生成結果偏黃嘅問題，我有跑幾張對比圖作參考。

3，模型複雜度降低問題。理論上及實測，v5比之前版本更乾淨同多元化，在部分提示詞作用下能有更精準表現。我亦做咗幾張對比圖比較。呢次訓練集冇用過度複雜素材，因我認為過度複雜圖像易令結果過擬合，必然導致細節缺失。

目的：我希望得到一個同之前版本有大差別嘅模型，而唔係發布幾乎一模一樣嘅模型。大家嘅反饋係好好嘅試錯機會，單靠自己冇咩試錯成本。喺下個版本我會努力增加不同畫風嘅素材量，令唔同素材嘅畫風能好好融合及分離，用特定prompt切換畫風，可能要用新嘅訓練技巧。多謝大家嘅反饋！

Summarize some issues regarding the v5 version.

1, Lora compatibility and issues with limbs and blurred eyes. Lora compatibility is that I used too much final weight for this training, and in some cases, overfitting may occur. This optimized version is the one that reduces the corresponding weight, and the limb collapse rate and compatibility with some Loras should be better. I have run several comparison charts of Loras trained with v4.1 for reference. The problem of blurred eyes should be the reason why I trained style_1. The eyes in the original material used are blurry, and can be improved by using style_3 or 4.

2. Exposure issues with volume light. I did not encounter this issue during testing, and the reason for it should be that I used the noise offset training parameter to increase the sensitivity of the model to light related prompt words, resulting in brighter results when the same weight of light prompt words were used. I suggest trying not to use parentheses and numbers to increase the weight. Due to the sensitivity of sdxl to prompt words, you can try repeating the same prompt words multiple times to avoid extreme results. At the same time, using this parameter is to fix the problem of generating yellow results under a small number of prompt words. I have run several comparison graphs for reference.

3. The problem of reduced model complexity. In theory and in practice. V5 should be a cleaner and more diverse model than the previous version, and with the help of some prompts, it should be able to achieve more accurate performance. Similarly, I ran several comparison charts for comparison. This training set did not use overly complex materials because I believe that overly complex images tend to overfit the results, which inevitably leads to a certain degree of detail loss.

Purpose: I hope to obtain a model that is significantly different from the previous version, rather than releasing a model that is almost identical to the previous version. This feedback from everyone is a great opportunity for trial and error, and I really don't have any trial and error costs on my own. In the next version, I will try to increase the amount of materials for different art styles, so that the art styles of different materials can be well integrated and separated. Using specific prompts to switch art styles may require some new training techniques. Thank you for your feedback!