LoRA Epoch 4 vs Epoch 6 Look the Same? img2img vs txt2img

LoRA epoch 4 and epoch 6 drawing the same picture? Then the LoRA is not being applied, and img2img versus txt2img is the place to check. I'm Labo, the lighthouse keeper. Today: why my LoRA epoch 4 and epoch 6 files drew the same picture, and how that taught me the difference between img2img and txt2img the hard way. What I learned: if different epochs give the same result, the LoRA isn't working at all.

LoRA epoch 4 and epoch 6 gave the same picture, img2img vs txt2img, Labo and the student comparing two identical prints

Two files with different names, two pictures exactly alike.

Last time we wired the ComfyUI default workflow, and I mentioned getting lost over one box, denoise. The student, whose dad is a Korean webtoon artist with no days off, is learning this to help him draw his manga with AI at home. She sat down with last time's question.

"What are epoch 4 and epoch 6?"

Student: Labo, what's an epoch? And why is "the same" bad?

Labo: One epoch is one full lap through the training images. Take my second style LoRA, trained June 5:

  • Where: an RTX 4090 rented on RunPod
  • Images: 55
  • Repeats: 5 (the number at the front of the folder name, like 5_lbtoon01)
  • Epochs: 10
  • Total: 55 × 5 × 10 = 2,750 steps
  • Captions: cross-checked by Claude and GPT against each other, about 95% accurate

Student: So 55 times 5 is one lap, and it ran ten laps?

Labo: Right: 275 steps a lap, ten laps, 2,750. At the end of each lap, the LoRA as it is right then can be saved as a file, an epoch file. This run gave me epochs 1 to 9 plus the final one, names ending in a number, like lbtoon01-000004.safetensors for epoch 4. How often to save is a training setting.

"Why so many files? Isn't the last one the best?"

Labo: Not always. Stop too early and it hasn't learned enough; run too long and it copies the training images so closely it goes stiff. So you keep several and try each.

Student: Which one won for you?

Labo: For this LoRA, epoch 5 looked best, and I used 6 as well. Not the final file.

Student: Then epoch 4 and 6 should come out a little different. One has learned more.

Labo: Exactly. Hold on to that; it's today's key.

"I put in a color palette. Why won't the style show up?"

Labo: June 6. I put a color palette picture in as the starting point and tried to lay the LoRA's style over it: img2img.

Student: Image to image?

Labo: Yes. Last time KSampler got an empty latent, a blank canvas, and drew from the prompt alone: txt2img. In img2img, your own picture turned into a latent goes in instead, and denoise decides how much gets redrawn. Low keeps most of it; 1.0 almost starts over.

Student: So if the style doesn't show, turn the number up.

Labo: That's what I thought. I ran denoise from 0.4 all the way to 1.0, even 0.95. The style didn't show anywhere.

Student: Even at 1.0? You said that almost starts over.

Labo: Which confused me more. I kept blaming the number.

img2img vs txt2img difference in ComfyUI, blank latent start vs your picture as the start, what denoise means

A blank start is txt2img; your picture as the start is img2img. Denoise is how much of the start gets erased.

"If different epochs give the same picture, is the LoRA not working?"

Labo: Then I swapped LoRA files. One picture with epoch 4, one with epoch 6. The two came out exactly the same.

Student: What I said earlier. If one learned more, they have to differ, even a little.

Labo: Yes. If different epochs give the same result, the LoRA isn't working at all. It was attached, but it had no effect on the picture.

Student: So no denoise number would have helped...

Labo: None. I thought "the style is weak" when it was "the style isn't getting in at all." A different-sized problem.

You can use this as a check: leave everything else alone, swap only the epoch file, make two pictures. Anything changes, the LoRA is at work. Identical, it's cut off somewhere. Where my img2img graph broke that day isn't in my notes. Changing the method fixed it.

LoRA not working check, swap only the epoch file, epoch 4 and epoch 6 identical pictures mean the LoRA isn't applied

Same picture from two epochs: check the line before the sail.

"So how did you fix it?"

Labo: I stopped feeding the palette in as a latent, and split the job in three:

  • 1. Back to txt2img. A blank latent into KSampler again, denoise 1.0, drawing from scratch.
  • 2. Structure held by ControlNet. A blank start lets the layout go anywhere, so shape and composition went to ControlNet, at strength 1.0 in that June 6 txt2img graph.
  • 3. Color on its own. I made color separately in a "color mode" pass and merged the two with the same seed.

Student: Why does the same seed help?

Labo: The seed numbers the random pattern a picture starts from. Same graph, size and seed, same start, so the structure and color pictures line up.

Student: It didn't work all at once, so you split it.

Labo: One job per step, and it came loose.

Student: So now the style comes out. Can I draw the same girl in different outfits, panel after panel?

Labo: That's exactly where I got stuck next. A consistent character across panels sounds simple. Next time: same LoRA, same character, but put her in different clothes and she looked like a different person.

FAQ

Q. What is a LoRA epoch file?
The LoRA saved at the end of each lap through the training images. From 55 images × 5 repeats × 10 epochs I got epochs 1 to 9 plus the final file; epoch 5 was best.

Q. My epoch 4 and epoch 6 pictures are identical. Why?
If only the LoRA file changed and nothing differs, the LoRA isn't working. It happened to me in img2img; going back to txt2img fixed it.

Q. What's the difference between txt2img and img2img?
txt2img starts from a blank latent; img2img from your own picture. Denoise is how much of that start gets redrawn.

What the artist did: ran a palette img2img from denoise 0.4 to 1.0 (0.95 included), swapped epoch 4 and epoch 6 and noticed the pictures were identical, then split the job: txt2img, structure and color made separately and merged with the same seed.
What the AI did: Claude and GPT cross-checked the captions (about 95%); a RunPod 4090 ran 2,750 steps on 55 images; the picture was drawn from a blank latent at denoise 1.0, with ControlNet (1.0) holding the layout and a "color mode" pass doing the color.

Official pages: RunPod (paid; check hourly prices on the site), ComfyUI on GitHub (open source, free), the original ControlNet project on GitHub (open source, free), Claude (Anthropic) and ChatGPT (OpenAI, the "GPT" above). GitHub is a site where code is shared in the open, so anyone can see what's inside before running it.

Written October 2026 from Labo's own records; file and trigger names are made-up examples. Node names and training screens may change between versions. The original Korean post: 로라 에폭 4랑 6 그림이 왜 똑같아요? (Korean)

Lighthouse log. If the epochs differ and the picture doesn't, the LoRA isn't working. I thought the wind was weak and kept trimming the sail; the line had come untied. When turning the wheel doesn't move the ship, check the line before the wheel.

Comments

Browse by topic

Latest posts

    See every post

    Popular posts from this blog

    First Kohya LoRA Done, Style Missing? Checking It in SwarmUI

    Draw Manga with AI at Home: LoRA, ComfyUI and ControlNet Map

    How Many Drawings to Train a LoRA? Picks, Captions, Trigger