Gemma TIPO Image 30K โ€” release and limitations / ๋ฐฐํฌ ์•ˆ๋‚ด

ํ•œ๊ตญ์–ด README ยท English README ยท Q4 installation

ํ•œ๊ตญ์–ด

์ด๋ฏธ์ง€์—์„œ ํ”„๋กฌํ”„ํŠธ์— ์“ธ ํƒœ๊ทธ๋ฅผ ์–ป๊ธฐ ์œ„ํ•œ ์‹คํ—˜์šฉ LoRA์ž…๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ๊ฐ€ ์ทจํ–ฅ์— ๋งž์ง€ ์•Š๊ฑฐ๋‚˜ ์ด๋ฏธ์ง€ยท์ง€์‹œ์™€ ๋‹ค๋ฅผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ž๋™ ์ •๋‹ต ํƒœ๊ฑฐ, ์บ๋ฆญํ„ฐ/์ž‘๊ฐ€ ์‹๋ณ„๊ธฐ, ์™„์„ฑ๋œ ํ•œ๊ตญ์–ด ํŽธ์ง‘ ๋ชจ๋ธ๋กœ ๋ณด์žฅํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ƒ์„ฑ ์ „์— ๊ฒฐ๊ณผ๋ฅผ ํ™•์ธํ•˜๊ณ  ์ˆ˜์ •ํ•˜์„ธ์š”.

  • ๋ฐฐํฌ ๋ชจ๋ธ: raspie/gemma4-tipo-lora-image-30k
  • ์„ ํƒ ์ด๋ฆ„: tipo_lora=image-30k, ์•ฝ 45.3MiB. ์ฒ˜์Œ ์„ ํƒํ•˜์—ฌ ์‹คํ–‰ํ•  ๋•Œ ๋‹ค์šด๋กœ๋“œํ•ฉ๋‹ˆ๋‹ค.
  • ํ•™์Šต: ์ด๋ฏธ์ง€ ๊ธฐ๋ณธ ์‚ฌ๋ก€ 30,000์žฅ + ๋ณด์กฐ 3,000ํ–‰, ์ด 33,000ํ–‰, 1ํšŒ ํ•™์Šต. ์–ธ์–ด ๋ถ€๋ถ„ rank 8 / alpha 16 LoRA์ด๋ฉฐ ๋น„์ „ยทPLEยทํ† ํฐ ์ž„๋ฒ ๋”ฉ์€ ๊ณ ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค.
  • ๊ธฐ๋ฐ˜: p-e-w/gemma-4-E2B-it-heretic-ara์˜ ๊ณ ์ • ํŒ๋ณธ c9a1d4c031981f14d86eeb0c7d87de7fafd34513. Google ์›๋ณธ์˜ Heretic ARA ํŒŒ์ƒ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค.
  • ์• ๋‹ˆยท์ผ๋Ÿฌ์ŠคํŠธ ๋ฐ ์Šคํƒ€์ผํ™”ํ•œ ๊ทธ๋ž˜ํ”ฝ์„ ์œ„ํ•œ ์ด๋ฏธ์ง€ ํƒœ๊ทธ ๋ณด๊ฐ•์— ์ดˆ์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ž์—ฐ์–ด๋Š” ์›๋ณธ Gemma์— ๋งก๊ธฐ๋Š” ์„ค์ •์„ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค.

์„ค์น˜์™€ ๊ถŒ์žฅ ์„ค์ •

  1. ์ €์žฅ์†Œ๋ฅผ ์—…๋ฐ์ดํŠธํ•˜๊ณ  ComfyUI์˜ Python์œผ๋กœ requirements.txt๋ฅผ ์„ค์น˜ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด ํ๋ฅผ ๋งˆ์นœ ๋’ค ComfyUI๋ฅผ ์žฌ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. Windows/Linux ๋ช…๋ น.
  2. Gemma Prompt Expander Native์—์„œ **model_source=heretic_ara_q4**๋ฅผ ์„ ํƒํ•˜๊ณ  clip์„ ์—ฐ๊ฒฐํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ํŒŒ์ผ ๊ฒฝ๋กœ๋Š” ๋น„์›Œ๋‘๋ฉด ์ž๋™์œผ๋กœ ๋‹ค์šด๋กœ๋“œํ•ฉ๋‹ˆ๋‹ค.
  3. **tipo_lora=image-30k**๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ๋ณ„๋„ LoRA ๋กœ๋”๋Š” ํ•„์š” ์—†์Šต๋‹ˆ๋‹ค.
  4. ์šฐ์„  **output_format=tags**๋กœ ์ด๋ฏธ์ง€์— ๊ด€๋ จ๋œ ํƒœ๊ทธ๋ฅผ ํ™•์ธํ•˜์„ธ์š”. ์ž์—ฐ์–ด๋„ ์›ํ•˜๋ฉด output_format=both, **tipo_lora_scope=tags_only**๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ฒฝ์šฐ ์›๋ณธ Gemma๊ฐ€ ์žฅ๋ฉด์„ ๋ฌธ์žฅ์œผ๋กœ ์“ฐ๊ณ , LoRA๊ฐ€ ๊ทธ ์žฅ๋ฉด์„ ํƒœ๊ทธ๋กœ ํ‘œํ˜„ํ•ฉ๋‹ˆ๋‹ค.
  5. ํ‰๊ฐ€ ์‹œ์ž‘ ์„ค์ •์€ temperature=0.5, repetition_penalty=1.15, max_tokens=512, seed=42์ž…๋‹ˆ๋‹ค. ์ทจํ–ฅ์— ๋งž๋Š” ๊ฒฐ๊ณผ๋ฅผ ๋ณด์žฅํ•˜๋Š” ์„ค์ •์€ ์•„๋‹™๋‹ˆ๋‹ค.

๊ธฐ์กด ์›Œํฌํ”Œ๋กœ์˜ ๊ธฐ๋ณธ๊ฐ’ tipo_lora=none, model_source=connected_clip์€ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. tipo_lora_scope๋„ ๊ธฐ์กด ๊ธฐ๋ณธ๊ฐ’ both๋ฅผ ์œ ์ง€ํ•˜๋ฏ€๋กœ ์œ„ ์„ค์ •์„ ์ง์ ‘ ์„ ํƒํ•˜์„ธ์š”. Gemma TIPO Native์˜ ํƒœ๊ทธ ์ „์šฉ ๋…ธ๋“œ์—์„œ๋„ image-30k๋ฅผ ์„ ํƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฏธ ๋ณ‘ํ•ฉํ•œ ๋ชจ๋ธ ์œ„์— ๊ฐ™์€ LoRA๋ฅผ ์ค‘๋ณต ์ ์šฉํ•˜์ง€ ๋งˆ์„ธ์š”. ์‹œํ—˜ํ•˜๋ ค๋ฉด ์ด๋ฏธ์ง€ ํƒœ๊ทธ API ์˜ˆ์ œ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. API ์š”์ฒญ JSON์ด๋ฉฐ ์บ”๋ฒ„์Šค ์›Œํฌํ”Œ๋กœ ํŒŒ์ผ์€ ์•„๋‹™๋‹ˆ๋‹ค. LoadImage์˜ ํŒŒ์ผ๋ช…์„ ๋ฐ”๊พธ์„ธ์š”.

์‹คํ–‰ ์‚ฌ์–‘๊ณผ ๋ถ€๋‹ด

๊ฐ€๋ฒผ์šด CPU์šฉ ํ”Œ๋Ÿฌ๊ทธ์ธ์ด ์•„๋‹™๋‹ˆ๋‹ค. LoRA ํŒŒ์ผ์ด ์ž‘์•„๋„ Gemma ์ „์ฒด ์ถ”๋ก ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.

ํ•ญ๋ชฉ ์•ˆ๋‚ด
GPU ๊ฒ€์ฆ ์žฅ๋น„๋Š” RTX 4070 Ti SUPER / RTX 5060 Ti, ๊ฐ๊ฐ 16GB VRAM. ์—ฌ์œ  ์žˆ๊ฒŒ ์“ฐ๋ ค๋ฉด ์ด ๊ธ‰์˜ NVIDIA CUDA GPU๋ฅผ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค. ์ตœ์†Œ ์‚ฌ์–‘์„ ๊ฒ€์ฆํ•œ ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค.
RAM 32GB ์ด์ƒ์„ ์—ฌ์œ  ์žˆ๋Š” ์šด์˜ ๊ธฐ์ค€์œผ๋กœ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค. ์ธก์ •ํ•œ ์ตœ์†Œ ์š”๊ตฌ๋Ÿ‰์€ ์•„๋‹™๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ์ƒ์„ฑ ๋ชจ๋ธยทVRS์™€ ํ•จ๊ป˜ ์“ฐ๋ฉด ๋” ํ•„์š”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Q4 ๋‹ค์šด๋กœ๋“œ ํ…์ŠคํŠธ+ํ† ํฌ๋‚˜์ด์ € ์•ฝ 3.22GiB, ์ด๋ฏธ์ง€ ์‚ฌ์šฉ ์‹œ ๋น„์ „ ํฌํ•จ ์•ฝ 4.14GiB. LoRA ์•ฝ 45.3MiB ์ถ”๊ฐ€. ๋””์Šคํฌ ํฌ๊ธฐ์ด๋ฉฐ VRAM ์š”๊ตฌ๋Ÿ‰์ด ์•„๋‹™๋‹ˆ๋‹ค.
๋Œ€์ฒด ๋ฒ ์ด์Šค ์—ฐ๊ฒฐํ˜• BF16 ๋„ค์ดํ‹ฐ๋ธŒ ๋ชจ๋ธ์€ ์•ฝ 9.57GiB. Q4 ๊ฒฝ๋กœ์—์„œ๋Š” ๋ณ„๋„๋กœ ๋ฐ›์„ ํ•„์š”๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค.
์‹œ๊ฐ„ Linux / 5060 Ti์˜ 24์žฅ ํƒœ๊ทธ ํ‰๊ฐ€์—์„œ ๋…ธ๋“œ ํ˜ธ์ถœ ์‹œ๊ฐ„ ์ค‘์•™๊ฐ’ ์•ฝ 36์ดˆ, ๋ฒ”์œ„ ์•ฝ 16~126์ดˆ. ์ถœ๋ ฅ ๊ธธ์ด์™€ ์ ์žฌยท๋‹ค๋ฅธ ์ž‘์—…์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค. ์†๋„ ๋ณด์žฅ์ด ์•„๋‹™๋‹ˆ๋‹ค.
๋‘ ์ถœ๋ ฅ ์ž์—ฐ์–ด โ†’ ํƒœ๊ทธ ๋‘ ๋ฒˆ ์ƒ์„ฑํ•˜๋ฏ€๋กœ ํƒœ๊ทธ๋งŒ ์ƒ์„ฑํ•  ๋•Œ๋ณด๋‹ค ์ž‘์—…์ด ๋Š˜์–ด๋‚ฉ๋‹ˆ๋‹ค.

GPU ๋ฉ”๋ชจ๋ฆฌ๋Š” ์••์ถ• ๊ฐ€์ค‘์น˜ ์™ธ์—๋„ ๋ณต์› ๋ฒ„ํผยท์ด๋ฏธ์ง€ ์ฒ˜๋ฆฌยทKV ์บ์‹œยทLoRA๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฏธ์ง€ ์ƒ์„ฑ ๋ชจ๋ธ๊ณผ ๋™์‹œ์— ์˜ฌ๋ฆฌ๋ฉด OOM์ด๋‚˜ CPU ์˜คํ”„๋กœ๋”ฉ์œผ๋กœ ์ธํ•œ ์ง€์—ฐ์ด ์ƒ๊ธธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•œ ์ด๋ฏธ์ง€ยท์งง์€ ์ถœ๋ ฅ๋ถ€ํ„ฐ ์‹œ์ž‘ํ•˜๊ณ , ๋ฉ”๋ชจ๋ฆฌ๊ฐ€ ๋ถ€์กฑํ•˜๋ฉด ๋‹ค๋ฅธ ๋ชจ๋ธ์„ ํ•ด์ œํ•˜๊ฑฐ๋‚˜ ์ถœ๋ ฅ ๊ธธ์ด๋ฅผ ์ค„์ด์„ธ์š”. CPU ์ „์šฉยท8GB GPUยทAMD/AppleยทWindows ์‹คํ–‰์€ ์ด๋ฒˆ ๋ฐฐํฌ์˜ ์‹ค๊ธฐ ๊ฒ€์ฆ ๋ฒ”์œ„๊ฐ€ ์•„๋‹™๋‹ˆ๋‹ค. ์„ค์น˜ ๋ช…๋ น ์ œ๊ณต๊ณผ ํ•ด๋‹น ํ™˜๊ฒฝ์˜ ์‹คํ–‰ ๊ฒ€์ฆ์€ ๊ตฌ๋ถ„ํ•ฉ๋‹ˆ๋‹ค.

์‹ค์ œ ๊ฒฐ๊ณผ์™€ ๋‚จ์€ ํ•œ๊ณ„

ComfyUI Q4 ๊ฐœ๋ฐœ ์ด๋ฏธ์ง€ 24์žฅ์˜ ๋™์ผ ์กฐ๊ฑด ๋น„๊ต์ž…๋‹ˆ๋‹ค. ์ด๋ฏธ์ง€์™€ ์‚ฌ์šฉ์ž๊ฐ€ ์ค€ ์ด๋ฆ„์„ ์ž…๋ ฅํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฆ„์„ ์ด๋ฏธ์ง€์—์„œ ์•Œ์•„๋งžํžŒ ์‹œํ—˜์€ ์•„๋‹™๋‹ˆ๋‹ค.

ํ•ญ๋ชฉ ์›๋ณธ ์ด์ „ 3์ฒœ ์žฅ ๋น„๊ต ํ›„๋ณด ์ด๋ฒˆ 3๋งŒ ์žฅ
PixAI ํƒœ๊ทธ์™€์˜ ํ‰๊ท  F1 0.141 0.271 0.352
์ž…๋ ฅ ์ด๋ฆ„ ๋ˆ„๋ฝ / 93ํšŒ 10 17 2
๋นˆ ์ถœ๋ ฅ / 24์žฅ 0 0 0

F1์€ PixAI๊ฐ€ ๋งŒ๋“  ํƒœ๊ทธ์™€์˜ ์ผ์น˜ ์ง€ํ‘œ์ด๋ฉฐ 35.2%์˜ ์ธ๊ฐ„ ํ‰๊ฐ€ ์ •ํ™•๋„๋ฅผ ๋œปํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด์ „ 3์ฒœ ์žฅ ๋น„๊ต ํ›„๋ณด๋Š” ๋ณ„๋„ ๋‚ด๋ถ€ ์‹คํ—˜์ด๋ฉฐ ๊ณต๊ฐœ pilot-001/002์™€ ๋‹ค๋ฅธ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์ตœ์ข… ๋ถ„๋ฆฌ ํ‰๊ฐ€ ์ž๋ฃŒ๋Š” ์‚ฌ์šฉํ•˜์ง€ ์•Š์•˜๊ณ , ์ œํ•œ๋œ ๊ฐœ๋ฐœ ์ž๋ฃŒ์˜ ๊ฒฐ๊ณผ๋งŒ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค.

ํ•œ๊ตญ์–ด ์ค‘์‹ฌ ์š”์ฒญ 12๊ฑด์—์„œ๋Š” ํ•„์ˆ˜ ํƒœ๊ทธ ๋ˆ„๋ฝ์ด ์ด์ „ 3์ฒœ ์žฅ 17๊ฐœ โ†’ ์ด๋ฒˆ 18๊ฐœ, ์ง€์ • ์ด๋ฆ„ ๋ˆ„๋ฝ์€ 0๊ฐœ โ†’ 3๊ฐœ์˜€์Šต๋‹ˆ๋‹ค. VRS ์ฐธ๊ณ ๋ฅผ ์ค€ 6๊ฑด์—์„œ๋Š” ์ด๋ฆ„ ๋ˆ„๋ฝ 3๊ฐœ์™€ ์ถฉ๋Œ ํƒœ๊ทธ๋ฅผ ๋‚ด๋Š” ์š”์ฒญ 2๊ฑด์ด ๋‚จ์•˜์Šต๋‹ˆ๋‹ค. ๋‹จ์–ด ๋Œ€์กฐ๋Š” ๋™์˜์–ด๋„ ๋ˆ„๋ฝ์œผ๋กœ ์…€ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

์‹ค์ œ ์ถœ๋ ฅ์—๋Š” ์งง์€ ๋จธ๋ฆฌ๋ฅผ ๊ธด ๋จธ๋ฆฌ๋กœ ๋ณ€๊ฒฝ, ์—ฌ์„ฑ ํ•œ ๋ช…์— ๋‚จ์„ฑ ํƒœ๊ทธ ์ถ”๊ฐ€, ์ง€์ •ํ•œ ์ž‘๊ฐ€ยท์บ๋ฆญํ„ฐยท์ž‘ํ’ˆ๋ช… ๋ˆ„๋ฝ์ด ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. VRS๋ฅผ ์—ฐ๊ฒฐํ•ด๋„ ์ฐธ๊ณ  ์ด๋ฏธ์ง€์˜ ํŠน์ง•์ด ์„ž์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ban_tags์˜ ์ •ํ™•ํ•œ ๋ฌธ์ž์—ด ์ œ๊ฑฐ๋งŒ์œผ๋กœ ์˜๋ฏธ์ƒ ์ถฉ๋Œ์ด ํ•ด๊ฒฐ๋˜์ง€๋Š” ์•Š์Šต๋‹ˆ๋‹ค. ํ•™์Šต์ด ์„ฑ์ธ ํ‘œํ˜„์„ ํฌํ•จํ•˜๋Š” ํƒœ๊ทธ ์˜์—ญ์„ ๋‹ค๋ฃจ๋ฏ€๋กœ ์›์น˜ ์•Š๋Š” ํƒœ๊ทธ๊ฐ€ ๋‚˜์˜ฌ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.

๋‘ ์ถœ๋ ฅ ๋น„๊ต์—์„œ๋Š” LoRA๋ฅผ ํƒœ๊ทธ ๋‹จ๊ณ„์—๋งŒ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ž์—ฐ์–ด ๊ฐœ์„ ์€ ์ž…์ฆํ•˜์ง€ ์•Š์œผ๋ฉฐ ์›๋ณธ Gemma์˜ ์ž˜๋ชป๋œ ์ด๋ฏธ์ง€ ํ•ด์„์ด ํƒœ๊ทธ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฏธ์ง€ ํƒœ๊ทธ๊ฐ€ ๋ชฉ์ ์ด๋ฉด ์šฐ์„  ํƒœ๊ทธ ๋‹จ๋… ๋ชจ๋“œ๋ฅผ ๋น„๊ตํ•˜์„ธ์š”. ๋‹จ๋ณด๋ฃจ ์ค‘์‹ฌ ์ž๋ฃŒ๋กœ ํ•™์Šตํ–ˆ์œผ๋ฏ€๋กœ e621 ์ „์ฒด ์–ดํœ˜ยท๋‹ค๊ตญ์–ด ๋ฒˆ์—ญ ์™„์ „์„ฑ๋„ ๋ณด์žฅํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.

๊ฐœ์ธ ์ด๋ฏธ์ง€ยท์›๋ณธ ํƒœ๊ทธยทํ”„๋กฌํ”„ํŠธยทVRS ๊ฒฝํ—˜ยทํ‰๊ฐ€ ์›๋ฌธ์€ ๊ณต๊ฐœํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. HF์—๋Š” ์–ด๋Œ‘ํ„ฐยท์„ค์ •ยท๋ฌธ์„œยท์ง‘๊ณ„๋œ ํ•™์Šต/ํ‰๊ฐ€ ์š”์•ฝ๋งŒ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ž๋™ ์ถ”๊ฐ€ ํ•™์Šต์€ ์—†์Šต๋‹ˆ๋‹ค. ๋งˆ์Œ์— ๋“ค์ง€ ์•Š์œผ๋ฉด **tipo_lora=none**์œผ๋กœ ์›๋ณธ ์ถœ๋ ฅ์„ ์‚ฌ์šฉํ•˜๊ฑฐ๋‚˜ CSV ๋…ธ๋“œ๋ฅผ ์‚ฌ์šฉํ•˜์„ธ์š”.

English

Image 30K is an experimental image-tag adapter. Results may be incorrect or simply not match your taste. Review and edit its output before image generation. It is not a reliable artist/character identifier or a finished instruction-editing model.

Download: raspie/gemma4-tipo-lora-image-30k. The ~45.3MiB adapter is fetched on first execution with tipo_lora=image-30k. It was trained once on 30,000 base image cases + 3,000 supplementary rows; rank 8 / alpha 16 language LoRA, with frozen vision, PLE and embeddings. Use the exact Heretic ARA base revision above, not an already merged adapter checkpoint.

Update this node pack, install requirements.txt with ComfyUI's Python, and restart after your queue finishes. Choose model_source=heretic_ara_q4, disconnect clip, and leave model paths empty for automatic, hash-checked downloads. No extra LoRA loader, llama.cpp, PEFT or Transformers is needed for this inference path. See Q4 setup.

Start with output_format=tags for image tags. For both outputs, select output_format=both, tipo_lora_scope=tags_only: base Gemma writes prose, the adapter renders that scene as tags. Existing workflow defaults remain unchanged; select these settings explicitly. The tag-only Gemma TIPO Native node also supports this adapter. Evaluation settings were temperature 0.5, repetition penalty 1.15, max tokens 512 and seed 42. The image API example requires your own input image; it is a POST /prompt body, not a canvas workflow. Offline adapter files belong in ComfyUI/models/loras/rs_gemma_tipo/image-30k/; both files must match the pinned hashes.

This is computationally demanding despite the small adapter. Tested GPUs were 16GB RTX 4070 Ti SUPER and RTX 5060 Ti under Linux. We recommend that class of NVIDIA CUDA GPU and 32GB+ system RAM for headroom, not as verified minimum requirements. The Q4 text/tokenizer download is ~3.22GiB, or ~4.14GiB with vision, plus the adapter. The alternative connected BF16 base is ~9.57GiB and is unnecessary for the Q4 path. Disk sizes are not VRAM budgets. Image processing, dequantization buffers, KV cache, LoRA, VRS and other generation models need additional memory. OOM/offloading can cause failures or substantial delays. Start with one image and shorter output; unload other models when needed. CPU-only, 8GB cards, AMD/Apple and Windows runtime were not validated.

In the 24-image 5060 Ti development run, node call time was about 36 seconds median, 16โ€“126 seconds range. These measured calls include variation in output/loading; they are not throughput promises. The dual-output mode runs two sequential generations.

On those 24 images, macro F1 against machine-produced PixAI tags was base 0.141, previous internal 3k candidate 0.271, Image 30K 0.352. Missing supplied names were 10 / 17 / 2 out of 93 occurrences respectively; all models returned nonempty outputs. Names were supplied, not recognized from pixels. Teacher agreement is not human accuracy; the 3k candidate is not public pilot-001/002.

Korean editing/VRS checks did not establish general improvement over the 3k candidate. Across 12 context requests, missing required tags changed 17โ†’18 and missing specified names 0โ†’3. In six retrieved-reference cases, the new model still omitted three names and emitted conflicting tags in two requests. Lexical checks can count synonyms as misses. Observed failures include shortโ†’long hair, an extra male subject and lost names. Unwanted adult-related tags are possible. Exact banned-tag removal cannot fix all semantic contradictions. VRS references may contaminate appearance. Prose was generated by the same base in this comparison, so no prose-quality improvement is claimed. Danbooru-focused training does not establish complete e621 or multilingual coverage.

This is a development-only evaluation; the reserved final set remains unused. Public files contain adapter weights, config, documentation and aggregate metrics only, with no source images, personal prompts, VRS experience or raw evaluation answers. No further training runs automatically. Select tipo_lora=none to return to the base, or use the model-free seasonal CSV node if the compute cost or outputs are not worthwhile.

Artifact provenance

The adapter SHA-256 is 9cfae7497b9031923b238e7e94afd1a654cc9e86d732d39007fbf6fc093bf4dc. training_summary.json contains sanitized aggregate evidence. Adapter tensor bytes are unchanged; config replaces a local path with the public base and exact revision. See LICENSE and NOTICE for Apache-2.0 and upstream attribution.

Downloads last month
24
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for raspie/gemma4-tipo-lora-image-30k

Adapter
(5)
this model