You've got 12 articles coming down ComfyUI, comets, tunes, errors, drawings, hangers, gotrolNet, take the Decth control space, and lock your face with IPadapter. There's almost nothing to stop you from doing this。
But there's one problem. I guess you hit it alreadyIt's coming out. It's coming outI don't know. Especially when we're users of 4060 8G, want to come up with a big 2048 x 2048? You'll see the D.C. Minutes, either directly out of memory or run like an old cow。
THE FIRST TIME I STEPPED ON THE PIT WAS PARTICULARLY TYPICAL: THERE WAS A HAPPY IMAGE OF A MAN WHO WANTED TO MAKE A CIRCLE OF FRIENDS ON THE COVER, WITH A MAGNIFYING FACE FULL OF MOSAICS AND HAIR. I WAS STUPID ENOUGH TO GO BACK TO THE SAMPLE, PLUS THE C.F.G., TO SPEND THE NIGHT..Because it's wrongI don't know. The figure and zoom in are two different things, and after the small picture is produced, the specific "magnification" step should be given. That's what this is about。
ONE, AI, ZOOM IN
One sentence:AI AMPLIFICATION IS AN ALGORITHM THAT RESTORES THE LOW-RESOLUTION MAP TO A HIGH-RESOLUTION AND IS “CHARACTERIZED BY ADDITIONAL DETAIL”, NOT A SIMPLE STRETCH。THE AMPLIFICATION OF THE CELL PHONE IS BY PLUG (MEANS OF PIXELS PLUGGED IN THE PIXELS), AND THE AI AMPLIFICATION MODEL (ESRGAN SYSTEM) IS TRAINED AND HAS SEEN TENS OF THOUSANDS OF HIGH-LEVEL MAPS, KNOWING WHAT IT SHOULD LOOK LIKE, AND RECREATES THE REAL TEXTURE. THE SAME LITTLE 512, AI ZOOMS OUT MORE OFTEN THAN YOU'RE HARD-BREEDING 1024。
Remember:IT'S A LITTLE FIGURE, THEN IT'S AN AI ZOOM, AND IT'S CLEARER THAN IT'S A BIG PICTURE AND IT'S LESS VISIBLE. THAT'S ALL THE OLD GUYS' DEFAULT。
Zoom in to the two main genres
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
ii. Where to put the model, what to load
The magnification model throws this directory together, suffix ...pth(OLD ESRGAN) OR .safetensors(new model) all lay down directly:
Photo by CafyUI/Models/upscale_Models/
Go to OpenModelDB / Hugging Face / Civitai's upscalers, free of charge. Select the content type (multi-source cross-validation):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Trail experience:I went back to put a 4x model in it, and I couldn't get the bottom of it, thinking it was broken and it was empty again. Later it was discovered that the model had to be updated with the browser web page (F5) and then the bottom box was updated. It's the same as the one that put Lora in it. I've done it again。
And a reminder:Do not use the wrong model type。Take the amplifier to show the real person's picture, and come out with a strange plastic sense; instead, put the written model in a binary, and the line will go bad. The content selection model saves a lot of back-to-work。
Route I: Model amplification - hand hands node (fast)
The simplest chain, three core nodes:
Load Image
Upscale Image
How do you get the port
IMAGE output from Load Image i don't know Enter Port
• The output of Load Upscale Model by Upscale Image upscale_model Enter Port
• The output of Upscale Image
Step Operation:
1. Add Node: Right-click search Load Upscale Model。
2. Plus: Search Upscale Image (Using Model) (real ImageUpscaleWithModel, imageupscaling)。
3. Connect: a feed map, a feeding model, two different mouths。
4. Select ModelThis post is part of our special coverage Global Voices 2011。
5. OutputSave Image, Queue。
One sentence:A node feed, a node feeding model, both of which plugged into the Upscale Image node, and made a big picture。the multiple is determined by the model - the 4x model transforms 512 into 2048, no change. it's a chain that doesn't spreadIt's fast, it's low, but it's not full of detailsIt's a "clear paste." We need to get more details on course two。
Clean-up ring: 10 minutes, zoom in to a big, clear picture (do it now)
LET'S START WITH THIS ONE. AFTER YOU'VE GOT AN EXTRA BIG MAP OF "CLEAN FROM THE PAST," YOU CAN CONFIRM FOR YOURSELF THE DIFFERENCE BETWEEN THE AMPLIFICATION OF AI AND THE ORDINARY STRETCH。
Preconditions: You have been able to come up with a map of a 512 or so. No, just one。
Target: Magnify the 512 small image with a model to 2048 large, high-cleansing, and to read the small image more clearly than "AI Zoom Vs Directly"。
Step (sight, don't jump):
1. Down Model: 4x-UltraSharp (general), drop it in ComfyUI/models/upscale_models/, flatten the condom folder。
2. Refresh Page:F5 REFRESH TO RECOGNIZE THE NEW MODEL IN THE DROP-DOWN BOX。
3. Add Node: Double-click blank search Load Upscale Model return car; Upscale Image (Using Model) return car。
4. Connect: IMAGE → Upscale Image mouth of VAE Decode (or Load Image) and Load Upscale Model output Upscale Image Upscale_Model mouth。
5. Select Model: Load Upscale Model choose 4x-UltraSharp。
6. Catch Output: Upscale Image Output Add Save Image。
7. OutputQueue Prompt。
Checkpoint1 output sized 2048 x 2048 (4x × 512), with edge texture clearly sharper; 2 dropbox not having your model = without refreshing or misplaying the directory, back to section ii; 3 report out of memory = the original figure is too large (if 1024 and ×4=4096), first down to the original figure or slashed in section vi。
Work (select, welcome comment area): The same map, one 4x-UltraSharp, one Real-ESRGAN, two contrasts — the difference between intuitive perception of “general sharpness” and “real nature” — the selection is not genre。
Route II: Proliferation amplification - by drawing quality, by supplementing details
The model zooms only "magnify what is available." If the map itself is not enough, it has to beSpread Zoom- Let the proliferation model be scaled up and redraw, and really add new details。
Most Common Ultimate SD Upscale(Defined node for "Ultimate SD Upscale" with ComfyUI Manager). It's smart in:cut the big picture into a piece of tile, run it one by one, img2img and spell it back seamlessly。AND EVEN IF IT'S 4K, IT'S JUST A "A PIECE OF TILE" AMOUNTTHAT'S THE KEY TO GETTING 4K OUT OF 4060 8GAnd run Flux on the 8G one idea: cut, block by block, and save。
The core is just a knobdenoise, press 0.3 ~0.5:
• Too low (<0.3): little detail, no running
• Sweet zone (0.3 ~ 0.5): structurally viable, but in real detail
• Too high (>0.5): The model begins to “play imagination” and the face changes, the image floats, equals a new picture。
Remember:the dnoise that spreads is the "how much" spin button, 0.3-0.5 is a life-saving zone, rather than a low than a high。
I love itFIRST ON ROUTE ONEPull up and clear the edgesAnd VAE Encode, go back to the low denoise KSamplerI don't know. It's equal to "Clean Zoom + Proliferation Complementation Details" in two steps, and it comes out big and solid. The chain is as follows (port-by-port check, note KSampler and several more ports):
Load Image VAE Encode
Port alarm:KSampler of this chain, apart from Upscale LatetAnd you'll also have to take model+CLIP Text Encode+VAE– Same as in the drawings, no one can run without them; both VAE Encode and VAE Decode use the same VAE。
KSampler Parameter: Denoise 0.35(0.3~0.4 most stable face protection), steps 20, cfg 6, sampler dpmpp_2m + karras, positive and negative hints followed in the original figure。
V. 4060 / 8G SPECIAL: NO-DISCLOSE 4K
8G OUT OF 4K, CORE SENTENCE:Don't zoom in on the whole. Let it cut and run。
· Preferred Path: Small scale (512/768) model zoom in to 2x (stop if needed) Larger than Ultimate SD Upscale slices。
· Chains are better than violence: 512 to 4K, not directly x 4, first 2x to 1024, then 2x to 2048, in two cleaner and less visible pressures。
· Cut Block Tool: Visible CuffyUI-TiledDiffusion Plugin, larger size, block by block, the same principle as Ultimate SD Upscale。
· Pure pixels: MAGNIFY ON TOP OF THE TOP WITH A DIRECT ROUTE ONE MODEL, SAVES THE MOST, AND 8G OR EVEN 6G CAN RUN。
I stepped through the pit and said it all at once: Go straight to the big picture 2048 → OOM; 2 dinoise, pull too high Up to 0.7 Face to another person single 4x hard crude edge, two times 2x better; 4 Model selected wrong class (a) A plastic sense of the person in the amplifier I forgot to refresh I thought it was broken again。
VI. QUIREMENTS OF PARLIAMENTS
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Common pits / mined areas
· It's empty: The web page/ error directory/ subfolder is not refreshed。
· It didn't get any clearer: Magnify with a pure model, but the original map is dead — pure magnification does not add to the details, and clearly moves to spread and magnify。
· The zoom in changed the picture: spread denoise too high, press back 0.3 ~ 0.4。
· OOM EXPLOSION: The original figure is too large / the single multiple is too high → + part。
· Plastic sense/fault: The magnification model does not match the content type, and selects the content。
ZOOM IN. YOUR STATIC MAP IS ALMOST COMPLETE. THE 14TH PART OF THE WORLD WE'RE MOVING FORWARD -- A LOT OF PEOPLE THINK THE 8G CARD CAN'T TOUCH THE VIDEO, ACTUALLY PLAY, AND I'LL SHOW YOU。