Flux Prompt Lanjutan LLM

This ComfyUI workflow le ve rag e s a modular, task-orie nte d archite cture to g e ne rate hig hly cohe sive and de taile d prompts for CLIP-L and t5xxl mode ls. The workflow e mploys thre e spe cialize d Larg e Lang uag e Mode l (LLM) module s, orche strate d in a se que ntial and inte rde pe nde nt manne r, to stre amline and optimize prompt cre ation.

Workflow Ove rvie w
Modul Analisis Input :

The workflow be g ins with a g e ne ral-purpose LLM re sponsible for parsing the input de scription.
It e xtracts se mantic me aning , ide ntifie s ke y visual and conte xtual e le me nts, and se parate s hig h-le ve l inte nt into two distinct pathways: CLIP-L Prompt and t5xxl Prompt g e ne ration.
CLIP-L Prompt Ge ne rator:

A se cond LLM module proce sse s the structure d input from the analysis phase to g e ne rate a concise , ke yword-drive n CLIP-L prompt.
This module prioritize s compactne ss and re le vance , e nsuring optimal compatibility with the CLIP-L mode l.
Output include s ke y compone nts such as main subje cts, gaya seni , se tting , lig hting , and color pale tte in a comma-se parate d format (e .g ., potret, photore alistic, sunse t, warm tone s, de taile d shadows).
t5xxl Prompt Ge ne rator:

Paralle l to the CLIP-L proce ss, a third LLM module produce s a richly de taile d, natural lang uag e de scription tailore d for t5xxl.
This module focuse s on g e ne rating up to 512 toke ns of de scriptive conte nt, cove ring aspe cts like :
Subje ct de tails (appe arance , pose , e xpre ssion, Pakaian ).
Environme ntal se tting s (time of day, archite ctural spe cifics, Properti).
Lig hting and color dynamics (inte nsity, kontras , harmoni ).
Sce ne composition (fore g round, middle g round, backg round e le me nts).
Atmosphe re and mood (e motive and symbolic nuance s).
Validasi dan Sinkronisasi:

Both outputs are validate d for se mantic and stylistic alig nme nt to e nsure consiste ncy be twe e n the CLIP-L and t5xxl prompts.
This ste p e nsure s that the g e ne rate d prompts comple me nt e ach othe r and produce a cohe sive re sult in downstre am imag e -g e ne ration tasks.

Ke y Fe ature s
Hie rarchical Prompt Eng ine e ring : Utilize s a multi-ste p, role -spe cific de sig n for modularity and pre cision.
Task-Orie nte d Workflow: Se parate s ke yword e xtraction (CLIP-L) from de taile d sce ne de scription (t5xxl) to optimize for mode l-spe cific stre ng ths.
Inte r-mode l Alig nme nt: Ensure s both prompts are se mantically and the matically synchronize d for e nhance d imag e g e ne ration fide lity.
Skalabilitas : The archite cture is adaptable for additional tasks, such as fine -tuning outputs for spe cific artistic style s or domains.