EclipseKww

Introduction(Character • Style • Multi-LoRA • Krea & Qwen notes)Hi friends! 🤗This is a friendly, experience-based walkthrough of how I build image LoRAs on Tensor.Art. I’m not a guru—just sharing what actually worked for me so you can skip a few potholes and have more fun. (This article is a merged compact version of previous three plus Flux Krea LoRA creation info~)We’ll cover:Character LoRA (Flux)Style LoRA (oil-painting vibe)Multi‑LoRA (several sub‑LoRAs inside one model)Flux Krea & Qwen LoRA (what’s the same, what’s different)Part 1. Character LoRA — “Skull Knight” (one of my early Flux projects) ⚔️💀I made a Flux character LoRA for a Skull Knight—not because the world needed it, but because I loved the vibe. It wasn’t my first-ever Flux LoRA, but it was one of the early ones and taught me a lot.ImagesTarget 15–40 solid images (I tried 12 for this case—doable, but more is safer).High quality + varied angles help a ton. Low-quality in → low-quality out (Flux is forgiving, but still).Core settings (Flux)Base: Flux.1 (default)Network: I like LoKr, but in project settings choose LyCORIS (LoKr selection has caused issues in some UIs; LyCORIS is the umbrella that supports LoKr/LoHA/LoCon, etc.).Trigger word: use a unique token (e.g., ek_sku11_kn1ght). I swap i/l with 1 to avoid collisions with normal words.Repeat / Epoch: e.g., 15 / 3 to save credits, then continue later if needed.Resolution: 1024×1024 is best for faces; 512×512 can still work for quick prototypes.Scheduler: cosine or cosine_with_restarts - 5% warming of total steps (constant/linear also fine).Optimizer: AdamW8bit (simple + memory‑friendly).Shuffle captions: ON. (OFF if # of image sets is low < 40)Keep N tokens: 1 (# of captions not to be shuffled).Noise offset: default 0.03 worked fine for me. (could be up to 0.05)conv_dim / conv_alpha (character): 4 / 1Labeling & promptsCaptioning: auto is fine; prepend the primary trigger to all captions (labeling tool helps).Add secondary descriptors at the end (e.g., “metal spikes”, “metallic surface”) to reinforce traits.Sample prompts during training should be simple so you can see the LoRA effect clearly.Training, picking epochs, publishingWatch epoch previews (4 images per epoch). Loss trends help, but your eyes win.If an epoch looks saturated (repeating) or artifacts creep in, stop early and save credits.Publish the best epoch. If you’re undecided, publish two versions (e.g., v1/v2 or pro/non‑pro) and let users pick.If you’re Pro, Continue Training from a good epoch is super handy.Result: even with 12 images at 512×512, I got a surprisingly usable LoRA. Flux does a lot of heavy lifting when your setup is sensible.Part 2. Style LoRA — Painting with Hopper 🎨This time I chased an Edward Hopper oil‑painting feel.Images30–100 images in a consistent style (I used 32 at 1024×1024).Add style‑specific caption hints like “flat colors”, “strong light–dark contrast”—these steer the vibe.Key differences from character LoRAStyle tends to learn fast → you often need fewer repeats/epochs.Example: repeat 10, epoch 3 was enough (pushing harder led to overfit or worse faces).conv_dim / conv_alpha (style): 8 / 2UNet LR: 0.0002 (default 0.0001 also okay; I nudged it for speed)Choosing the epochTreat it like tasting notes: Epoch 2 might be subtle and classy; Epoch 3 bold but riskier.A slightly higher-loss epoch can look better—trust visuals over numbers.Test strength: I usually try 0.8–1.0. That’s where most styles sing.Part 3. Multi‑LoRA — Several flavors in one 🍱Goal: train multiple sub‑LoRAs into a single “combo” model so you can call specific sub‑styles/characters by trigger.Folder idea:ComboChar/├── character_A/ (trigger: charA, 30 imgs)├── character_B/ (trigger: charB, 40 imgs)└── character_C/ (trigger: charC, 35 imgs)Caption pattern (at the start):ComboChar, charA or ComboChar, charB …Training switchesShuffle captions: TrueKeep N tokens: 2 (so those two triggers aren’t learned as content)After trainingUse ComboChar for the blended vibe, or ComboChar + charA to force a specific sub‑LoRA.During training, previewing all variants is limited; I check candidates after training, then re‑train if needed.I’ve used this trick for things like RPG Booster, Yellowstone/Yosemite, Winter Resort—packing variety into one model is surprisingly practical (and fun).Part 4. Flux Krea LoRA — Same recipe, new flavor 🍜Flux Krea popped up as the “new flavor” of Flux, so of course I had to try it with a Yor Forger LoRA.👉 The surprise? It’s basically Flux with a small twist.Network: LoRA only (no LoKr option here)network_dim / alpha: you must set it — a safe pick is 64 / 32 (or 48 / 24, 32 / 16)conv_dim / alpha: same story as Flux (character: 4/1, style: 8/2)Other knobs: repeats, epochs, LR scheduler, optimizer, captions → same playbook as FluxResults: clean, stable, training cost almost identicalSo if you already know how to make a Flux LoRA, you’ll feel right at home. Just remember: Krea is LoRA-only, so don’t forget to set network_dim / alpha.Part 5. Qwen Image LoRA — A slightly different spice ✅Next I tested the Qwen Image base model by building a Belleza LoRA. Honestly, the workflow still felt very familiar — but with a couple of quirks worth noting.Network options: LoRA and DoRA (I stuck with LoRA; DoRA is still new territory for me).network_dim / alpha: defaults to 32 / 32, but I had better luck with 32 / 16.conv_dim / alpha: default is 4 / 4; I dialed it down to 4 / 1 for character LoRA training.Training setup: same recipe as before. For Belleza I used repeat 20 / epoch 5, which trained smoothly. (# image data set = 32) For Momo and Yor Forger LoRAs, epoch 2 was good enough! So I stopped the training. How affordable the training for Qwen is! 🤗Results: just like Flux/Krea — stable, sharp, no extra credit cost.So the takeaway: Qwen doesn’t need you to learn anything brand-new. Think of it as Flux/Krea with slightly different defaults. Adjust network_dim/alpha and conv_dim/alpha to your taste, and you’re good to go.Part 6. Tiny cheat sheet (copy/paste)Character (Flux)conv_dim/alpha: 4/1Repeat/Epoch: ~15 / ~3 (scale to taste)Trigger: unique token at caption start (Keep N=1)Notes: watch visuals > loss; publish best epoch(s)Style (Flux)conv_dim/alpha: 8/2Images: 30–100 (consistent style, 1024×1024)Repeat/Epoch: low (e.g., 10 / 3)Add style hints in captions (“flat colors”, etc.)Multi‑LoRAStart of caption: MainTrigger, SubTriggerShuffle: True, Keep N tokens: 2Flux KreaNetwork: LoRA only → set network_dim/alpha = 64/32 (good default)conv_dim/alpha: same as Flux (character 4/1, style 8/2)Qwen ImageNetwork: LoRA → set network_dim/alpha = 32/16conv_dim/alpha: same as others (character 4/1, style 8/2)Epoch can be low for saturation. Check the progress closely and push the stop button!Part 7. ClosingLoRA making isn’t a strict science—it’s a creative recipe. Pick good ingredients (images), season with captions, adjust heat (params), and taste often (epoch previews). When it looks right, it is right.Have fun, share your models, and let the community riff on them. That feedback loop is where the real magic happens. ✨Good luck & happy training! 🤗

Simple tutorial to Wan 2.2 Video LoRA creationFirst of all — let me just say this. I was blown away by how good Wan 2.2 Video Model is. 🤯Honestly, I haven’t made that many videos before, but even I could tell right away: this model is awesome.Making a LoRA for Wan 2.2? Much easier than I expected. Generating videos with it? Even more fun. It feels almost too easy at times. So—welcome to Wan 2.2! In this post, I’ll share my first attempts at creating LoRAs and generating videos, mixing in my experiences with Momo, Yor Forger, and Belleza.Don’t expect a full textbook here. This is more of a friendly starter guide. Think of it as me handing over the starter settings that worked for me. And yes—I’m still a newbie myself. (So please, double check things before using them seriously 😅)Goal & SetupMy very first attempt was with Momo from the manga Dandadan. (If you’re curious, I also posted that LoRA separately!)Later I tried Yor Forger, and then Belleza. Each one taught me something new.Base model: Wan2.2 low-noise variant (like t2v-14B-low-noise)Output: short 3–8s anime-style clips (parks, cafés, jogging, lightsaber fight… you name it!)Data Prep — where the pain begins 🥲Honestly, preparing data takes way more effort than the actual training. Video LoRA is easier than still-image LoRA in many ways, but you do need good source clips.What I did:Around 50 images from 4 video clips (each clip ~5s long, recorded in 1080p 30fps). BTW, you can upload short clips directly (≤5s), but I didn’t. I preferred to control each frame manually.Extracted frames: roughly 1 frame every 8–10 frames → ~3–4 images per second. There can be several options to do this, but using linux command is the simplest way. You only need to install ffmpeg by "brew install ffmpeg" from the command prompt of your terminal. In my case, a Mac terminal.Then run:ffmpeg -i input.mp4 -vf fps=2 frames/frame_%04d.pngThis gives you ~2 frames per second.Resolution: I reduced them down to 960×544, all consistent. (Other ratios like 1:1 or 9:16 also work fine.)Cleanup: throw away cropped/blurred faces, and dedupe “almost identical” shots. (you’ll often end up with near-duplicates. That can cause overfitting, so I prefer handling images manually. Just my personal choice!)Captions? They say that video LoRA isn’t as sensitive to captions as still-image LoRA, but it’s still worth cleaning up. Auto-labeler worked fine for me. Just remember to stick your trigger word in front of captions later.Training Settings (LoRA)I wanted something that balances cost vs. quality. Video training is expensive!Base: Wan2.2-t2v-a14B-low-noiseNetwork: LoRATrigger: ek_momo_wan22_vl01 (pick your own trigger word, of course)Repeat/Epoch: 3 / 8 → total steps = images × 24. (Higher is better, but this was “good enough.”)UNet LR: 1.2e-4 — stable for me. (The default 1e-4 probably works too.)Scheduler: cosine (I like the smooth decay, but constant also works.)Warmup: 5% (important for cosine; for constant, basically 0).Optimizer: AdamW8bit (default, can’t be changed).LoRA Rank/Alpha: 64/32 (safe default). There’s a great tutorial on the TA Discord (guide channel + YouTube) if you want more depth.Target Frames / Frame Sample: 11 / 6.Save/Preview: every 2 epochs (saves credits). The cost of video Lora training is higher than the still image Lora's. I tried to reduce the training cost.Preview prompt (super simple):ek_momo_wan22_vl01, 1girl, medium shot, steady camera, soft daylight, clean background, smooth natural motion, subtle eye blink👉 Keep previews simple. If the prompt is too fancy, you won’t know whether the LoRA is working.Generation (Text → Video)Here’s where the real fun begins. A tiny tweak—steps, CFG, or LoRA weight—can change the entire video.My “sanity check” presetModel: Wan2.2-t2v-a14bRes / Len / FPS: 16:9 (480p), 3s, 16fps (Fast Mode)Steps / CFG: 8 / 1LoRA: 0.9–1.0 (depending on face sharpness)Showcase preset (720p)1280×720, 5–8s, 16fpsSteps / CFG: 25–30 / 4.2–4.6 (Quality Mode)LoRA: 1.0Prompt example (anime park scene):anime style, cel shading, clean lineart, flat colors, ek_momo_wan22_vl01, 1girl, medium shot, steady camera, sunlit park, blue sky, soft ambient light, smooth natural motion, gentle head turn, subtle eye blink, no scene cutsSwap out scene/background for cafés, jogging, or lightsaber duels.Urgent Update: Fast Mode ⚡Okay, here’s where it gets interesting.When I first tested Wan 2.2, Fast mode gave me pretty bad results. So I ignored it.But then I saw other users posting amazing videos generated in Fast mode, using only 8–9 steps. 🤯 Naturally, I retried with my Yor LoRA—and wow. The results were better than anything I got in Quality mode.So here’s the reality check:Fast mode = super cheap + super quickCan actually beat Quality mode if your LoRA is solidBut… sometimes resolution/detail is lower, so it’s not perfect for final showcaseStill, for experimenting and quick iterations? Fast mode is a game-changer.Closing ThoughtsSo far I’ve made three video LoRAs:Momo (lively, lots of movement—great first subject!)Yor Forger (less active dataset—reminded me that motion variety really matters)Belleza (beautiful test case—good for cinematic shots)Each one was a fun experiment, and I’m honestly still learning. But if I can get these results as a beginner, Wan 2.2 Video LoRA creation is much more approachable than I ever imagined.Fast mode, especially, opened my eyes. I’ll be using it a lot more going forward.Anyway—that’s my first experience. Hopefully this helps you get started too. 🤗