Ray Davis

[知识猫AI实验室 @GeekCatX]

提示词：

请生成一张高品质、真实拍摄质感的人物角色设定图（character sheet / character reference sheet）。

角色：{{角色��称}}
角色身份：{{角色基础身份}}
角色描述：{{角色描述}}
服装设定：{{服装描述}}
配饰设定：{{配饰描述}}
鞋靴设定：{{鞋靴描述}}
髮型设定：{{髮型描述}}

如有参考图：
- 仅使用参考图作为该角色脸部、骨骼结构、肤质与髮型的主要依据
- 保持年龄、种���、性别特征一致
- 脸部特征高度匹配参考图

输出要求：
- 纯白背景
- 16:9 横版
- 4K 解析度
- 布局清晰规范
- 各模块无重叠、无裁切
- 高精度细节，专业影视级角色卡风格

必须包含以下模块：

1. 头部细节组：
- 无配饰正视图
- 无配饰正侧视图
- 完整造型正视图
- 完整造型正侧视图

2. 全身正交转面组：
- 正面
- 左45度
- 正侧面
- 右45度
- 背面

3. 补充视角组：
- 半侧身视图
- 俯视顶视图
- 仰视底视图

4. 材质细节特写组：
- 服装面料特写
- 袖口 / 领口 / 腰部细节
- 鞋靴细节
- 配饰细节

一致性要求：
所有模块中的角色必须为同一人物，五官、髮型、肤色、身材、服装、鞋靴、配饰、光影、材质完全一致，不得出现角色漂移或视图差异。

[Ray Davis @rdombeki]

@simplifyinAI Used Airi a while back before i built mine

[Ray Davis @rdombeki]

@shinyflvres @simplifyinAI It works as expected of it, bugs will always be there on any product, vibecoded or not, nobody is perfect

[Ray Davis @rdombeki]

@shinyflvres @simplifyinAI Airi has been out for a while though

[🦋Hentai8cho🦋 @Hentai8cho]

Name: THE WITCHING HOUR (original)

Characters: Original Characters

Artist: Lewd Froggo

Generate AI videos and images on ourdream 🩷

https://t.co/CO7xzCMZjC

#hentai #porn #rule34 #rule34artist #uncensored #nude #chudai #erotic #hot #sexy #waifu

[Ray Davis @rdombeki]

@outsource_ q1??

[Ray Davis @rdombeki]

@AndrewCurran_ its not that bad, The AI hates corporations just like regular people

[ModelScope @ModelScope2022]

Only ~1.2B parameters active at a time. Edge0-8B-A1B-preview makes an 8B-class MoE practical for local inference.📜 Apache 2.0.
🤖 https://t.co/xJ8uHG80kB

⚡ Reaches 23.9–25.3 tokens/s with about 1.0 GiB peak active memory in the reported short-context benchmark.
🏆 Retains most of the FP16 base model’s quality, with an average gap of just 2.8 points across five benchmarks. MMLU-Pro rises from 65.8 to 70.1.
🧠 SSD expert offload keeps most weights outside active memory, while a prerouter predicts which experts each layer will need next.
🛠 Recover-LoRA offsets quality loss from 4-bit quantization and expert offloading. The current Preview runs locally through MLX on Apple Silicon.

[Ray Davis @rdombeki]

@queenmissyx the economy is hard
🤡

[𝖍𝖆𝖎𝖑𝖊𝖞 𝖖𝖚𝖊𝖊𝖓 🔥 @realhaileyqueen]

Are you old enough to watch this version of Elastigirl? 🔞

[Shegun 🌪️🌟 @ShegunTweets]

this show ended up being way better than i expected

[Ray Davis @rdombeki]

@mymaifu dp?

[David Hendrickson @TeksEdge]

🤯 So you can actually get a 100K context & ~50 tps on one 16GB GPU.

Putting Qwen3.8-27B onto an RTX 4070 Ti super is the kind of localmaxxing I ❤️ to hear about.

Recipe 👇
🧠 Qwen3.8-27B
🎮 RTX 4070 Ti SUPER — 16GB
📚 100,000-token context
⚡ 47–50 tok/s
💾 15.93GB VRAM

This recipe stacks several optimizations:
✧ 13.5GB hybrid GGUF
→ attention stays IQ4_XS
→ FFNs compressed to IQ3_S

✧ BeeLlama.cpp instead of stock llama.cpp
✧ KVarN compressed KV cache
→ K = kvarn5
→ V = kvarn4

✧ Last 1,024 KV tokens kept at higher precision
✧ MTP-2 speculative decoding

Changing the KV cache from kvarn5/kvarn5 → kvarn5/kvarn4 alone reportedly freed another ~6% VRAM and pushed context from 88K → 100K.

🎯 So this is what they did …
Weights → selectively quantized.
KV cache → selectively compressed.
Recent context → kept high precision.
Decode → speculated.

⚠️ This uses BeeLlama.cpp, a llama.cpp fork, not stock llama.cpp/LM Studio, and the 47–50 tps number is a community report.

Links 👇
🔗 Reddit r/LocalLLaMA/comments/1w1lq7u/qwen_38_27b_at_50_toks_with_100k_context_on_a/

[Ray Davis @rdombeki]

Hello world