ComfyUI from installation to mastery, 8G. 4060 Spectrum / Telegraph

In front of 15, you basically have ComfyUI I got it. I got it. I got it. I got it ComfyUI from installation to mastery, low-visibility scheme: Boogu let 8G run the big model that couldn't run You've tried that "breathless" Boogu model。

But there's a painting ceiling I haven't touched on purposeFLUX.1I don't know. It is generally recognized by those who play AI drawings that FLUX.1 shows people who are much more capable of layout than SDXL like skin sense, finger and Tulipi text. But it's full-precision 24G, and my 4060 is CUDA out of memory。

THAT'S WHAT I SAID. LOOK AT THE FLUX SHOW, FULL PRECISION, BIG MODELS, YOU'RE STUCK, YOU THINK YOU CAN'T USE IT. THEN IT TURNS OUT THE COMMUNITY MADE FLUX 1 QUANTIFIED VERSION OF GGF- PUT 23G ON 6-8G, 4060. THIS ONE TURNS IT FROM "SEEING PEOPLE PLAY" TO "I'M OUT."。

I. FLUX.1 WHAT IS IT

One sentence:FLUX.1 is a 12B parameter graphic model from Black Forest LabsIt's good to paint, to follow words, to point fingers, to draw words。

One of the most crucial things:THE FULL ACCURACY IS 24G, THE QUANTIFIED VERSION OF THE GGF PRESSES 6-8G, 8G, AND RUNS。

Remember:D.S.X.L. Level D.O. + Mass is only 8G card? FLUX.1 dev 's quantitative version of GGUUF is the answer。

II. Documentation preparation (by mistake)

GGUUF route four files, directory book (mistaken ComfyUI never clear):

CofyUI/models/unet/→flux1-dev-Q5_K_S.gguf
t t5xl_fp8_e4m3f.safetensors
CCfyUI/models/clip/clip_l.safetensors
CofyUI/models/vae/ae.safetensors

Two pits blocked earlier:
1. Must install ComfyUI-GGUUF plugins–Manager Search GGUF → ComfyUI-GGUF, restart. Normal Load Checkpoint cannot read .gguf files。
2. Text encoder with FP8 version t5xl(4.7G), don't use the fp16 version (9.4G) -- the latter eats almost 8G alone。

Trail experience:For the first time, FLUX, the text encoder went to his fp16 version (9.4G), thinking, "Quality is good," and the result was a Q5 main model with an 8G card, which was stuck to the OOM interface. Instead of fp8 (4.7G), it's just going down. On the low-profile card, the text encoder is often more explosive than the main model, and I stepped on it for you。

III. Hand-to-hand nodes (continuous)

FLUX does not default on the Load Checkpoint, using the GGUF special node. Nodal chain (port checked):

Unt Loader (GUF) →Model mouth of KSampler
DualCLIPLoader (type=flux)
VAE Loader (ae.safetensors)
EmttyLentImage (1024x1024) →latent_image mouth of KSampler
KSampler

One sentence:Spread the model to model mouth, double CLIP set flux through FluxGuidance into position, VAE to solve, EmttyLatent to set size。

Two critical pits (and 15 of Boogu's type=boogu co-source):
· The CFG of KSampler must set 1.0Not SDXL that set of 7-9. FLUX does not use the traditional CFG, which is set at the FluxGuidance node separately 3.5. The CFG set up high graphs to overexpose ash。
· Type of DualCLIPLoader must set flex, not sdxl, not sd3. sets the wrong image full of random numbers。

CYCLOPS: 10 MINUTES, A FLUX IMAGE / ELECTRON (DO NOW) IN 4060

THE PRINCIPLES OF THE PREVIOUS SECTIONS ARE TRUE. AFTER YOU'VE GOT ONE MORE TO USE, YOU CAN STILL CONFIRM THAT FLUX REALLY RUNS THROUGH YOUR 8G CARD。

Preconditions: Your ComfyUI is loaded + the ComfyUI-GGUUF plugin (four files according to section II). If you lose, go back to the second section and load the plugin + below the model。
Target: Run a FLUX.1 dev / Electrician product map (1024 x 1024) using 4060 (8G) to verify that "low-visibility storage can also produce FLUX-grade drawings"。

Step (sight, don't jump):
1. Down Model: CATALOGUE BASED ON FOUR DOCUMENTS IN SECTION II, F5 UPDATE PAGE。
2. Add Unt Loader (GGF): Double-click empty search unet loader (GGUUF) → unet_name select flux1-dev-Q5_K_S.gguf。
3. Plus DualCLIPLoader: Search clip for DualCLIPLoader → crip_name1 for t5xl_fp8_e4m3fn.safetensors, clip_name2 for clip_l.safetensors, type set flux。
4. Add VAE Loader: Search vae loader vae_name select ae.safetensors。
5. Encoding of hintsSearch clip text encode add CLIP Text Encode (Prompt) to write positive; copy one more positive negative (left empty, FLUX not required)。
6. Plus FluxGuidanceSearch for flex guidance plus flex guidance guidance guide strength set 3.5 → between the clit and KSampler mouth。
7. Add1024 wide, 1024 high, 1 batch to KSampler 's latent_image mouth。
8. Add KSampler: model for Unet Loader output, postive for FluxGuidance output, Negative for CLIP Text Encode, Vae for VAE Loader output; sampler=euler, Scheduler=simple, steps=20, cfg=1.0。
9. Add VAE Decode + Save Image: VAE Decode to KSampler output + VAE Loader output → Save Image to VAE Decode output。
10. Start ZoomCommand line python Main. py –lowvram start, point Queue Prompt out。

Checkpoint1 Normal: Node-by-Node Green, output a 10242 diagram, clean skin, normal fingers (FLUX strong), no pink / red; A abnormal A = a noise seizure / vae mouth of VA Decode = 8% of KSampler cfg = 1.0 + euler + simple; Aberrant B = pink / red: VAE error (no e. safetenensors or no VAE Decode = high vae) returns section 2 check; 4 Aberrant C= red error report” does not find a no no no point: common Load Checkpoint instead of Unet Loader (GUF), goes back to the CufyUI-GUF Plug; or VA Decode = high-altraffic D = low – 32, not started with OTM = too much = OTM = + 8R = high + = > > 8 > > > > > > > > > > > >。

Job (select): Move FuxGuidance from 3.5 to 5.0 and then out of one, compare the hints with the following; or replace "studio lighting" with "golden hour lighting" in the positive hint。

IV. CORE THREE EXCHANGE

my habits: daily brainless dev 20 step + guide 3.5; fast schnell 4 step (commercial); hintWrite long, fine, in natural languages, FLUX UNDERSTANDS THAT LONG DESCRIPTIONS ARE BETTER THAN STACK WEIGHTS。

Trail experience:For the first time, I used FLUX as SDXL, KSampler's cfg to fill seven and come out with a pile of dust and white trash. Half a day on FLUX's cfg has to be 1.0, leading strength to the FluxGuidance node. The new ones will step on it, remember。

V. Common pits at once

1. KSampler's cfg does not have 1.0 ~ Oversilent white, most common。
2. Normal Load / DualCLIPLoader • Output spun or miscalculated, GGUUF sets flux with type of UNnet Loader (GGUUF), CLIP。
3. fp16 t5xl (9.4G) under text encoder OOM, replace fp8 (4.7G)。
4. VAE WAS WRONG / NOT ANSWERED → Pink block, must be ae.safetensors and pick up VAE Decode's vae mouth。
5. - no. - lowvram 8G card OOM, start command plus –lowvram。
6. I wrote a negative hint → FLUX does not eat, does not write, does not work (empty negative node OK, but fills in words)。

VI. QUIREMENTS OF PARLIAMENTS

VII. 4060 / 8G SPECIALIZED & I STEPPED ON PITS

4060'S 8G RUNNING FLUXQ5_K_S IS DESSERT(~8G main model + fp8 t5 4.7G + VAE with –lowvram flow load within 8G). Points:

· MAIN MODEL Q5_K_S(~8G), LESS STABLE THAN Q4_K_S(~6.8G)。
· text encoder must fp8 version t5xl(4.7G), do not touch fp16 (9.4G)。
· start with -lowvram;slower but steady。
· Discrepancies lead, 10242I'm going to use the magnification of 13 more。
• Turning off other visible programs on the drawings。

Trail experience:Q8_0(13G), 8G CARD JUST DIED. BACK TO Q5_K_S SECONDS OUT. LOW VISIBILITY CARD Q5, QUALITY SUPPLEMENTED BY HINTS AND LATER MAGNIFICATIONS, DO NOT COMPETE WITH QUANTIFICATION。

FLUX WILL RUN, YOU'VE TOUCHED THE LOCAL PAINT CEILING. BUT YOU'RE GOING TO HAVE TO DO IT AGAIN AND AGAIN..How to efficiently assemble, update, sort out conflicts directly affects how many models you play behind。Title 17, Self-Defined Nodes + Manager, tells the story。