ArtificialGuyBR

Home / Blog

Technical notes on generative AI

Guides on training, prompting and running open models locally. Every claim is traced to a primary source — official docs, model cards, repositories, or the code itself.

97 articlesRSS

AI video (11)Training LoRAs (26)Models of 2026 (10)ComfyUI (16)Prompting and control (13)Fixing things (21)

Fix Automatic1111 Git Pull Break: Venv Reset and Rollback

Fix A1111 after a broken git pull by deleting venv — the official upstream path. Covers Forge repair, xformers speed gains, and what s…

Fixing things

RES4LYF Sampler Pack and DPM++ 2M vs 3M SDE Comparison

RES4LYF's 69 samplers vs built-in options, DPM++ 2M vs 3M SDE debate, and whether exotic high-stage samplers justify extra compute cos…

Illustrious and Anima Style Tags: What Actually Works

Verify Danbooru-style tags that Illustrious and Anima models actually recognize, test tag registration, and build reusable style short…

Models of 2026

AnimateDiff vs WAN 2.1: Motion Modules

AnimateDiff vs WAN 2.1: motion-module approach, where AnimateDiff wins, ComfyUI node guide, SDXL support.

Models of 2026

Automated LoRA Dataset Creation: Extract, Filter, Tag, QA

Build LoRA training sets from anime video and generative tools — automated extraction, tagging, and quality checks that keep the datas…

Training LoRAs

ComfyUI Setup: Measure Speed, Compare GPUs, Run on CPU

Measure ComfyUI iterations per second, compare RunPod GPU costs, and run on CPU when you have no GPU.

ComfyUI

CFG Scale in Stable Diffusion

How CFG scale controls prompt adherence in Stable Diffusion, optimal ranges per model family, and why distilled models use CFG 1.

Fixing things

IPAdapter FaceID vs LoRA for Character Consistency

Compare IPAdapter FaceID with trained LoRAs for character identity across images, including seed/pose control and expression preservat…

Training LoRAs

ReActor Face Swap in ComfyUI: Setup and Limits

Set up ReActor Fast Face Swap in ComfyUI — model files, restore-face and mask options, detailer chaining, and how swap differs from a…

ComfyUI

ComfyUI Film Production Pipeline: Consistency and Stitching

How filmmakers use ComfyUI for shot planning, character consistency across sequences, video compositing, and where the workflow breaks…

ComfyUI

ComfyUI Inpainting Workflows: A Practical Guide

A practical guide to ComfyUI inpainting workflows covering mask editing, ControlNet preservation, and masked-area sampling.

ComfyUI

ComfyUI Progress Bar and Day-One Settings

Add progress feedback to ComfyUI with Crystools and configure essential settings on first launch.

ComfyUI

ControlNet Shuffle Feature Blending Guide

ControlNet Shuffle for feature blending, Recolor for restoration, product/vehicle rendering from reference, and comics line work.

ComfyUI

ControlNet: Canny, Normal, and Panel Continuity

Use ControlNet for comics: Canny edge control, Normal-map 3D rotation, Reference style continuity, and stacking multiple control types.

ComfyUI

FLUX Consistent Characters Without Training

How FLUX achieves consistent character generation without LoRA training via reference conditioning, Redux adapters, and seed disciplin…

Training LoRAs

FLUX.1 FP16 vs FP8: Speed, Quality, and VRAM Trade-offs

Compare FLUX.1 Dev and Schnell in FP16 vs FP8 precision. Which mode fits 12GB vs 24GB GPUs, and where does image quality actually dive…

Fixing things

FLUX Prompt Terms for Eye Contact and Looking Away

Master FLUX gaze control with the right positive and negative prompts, IPAdapter tuning, and a reusable scaffold that works across mod…

FLUX Licensing in Plain Terms: A Practical Guide

What FLUX.1-schnell, dev, and pro each permit commercially — how LoRAs inherit restrictions and where the ambiguity remains.

Fixing things

Running FLUX on 8GB: fp8, GGUF, and what works

FLUX.1 on 8GB VRAM with fp8/GGUF quantization, CPU offloading, and schnell's 4-step budget. Sources on paths, files, speeds.

Fixing things

FLUX Photorealism Guide: Avoiding the Plastic Look in 2025

FLUX photorealism with prompt structure, CFG settings, skin texture, and LoRAs

Fixing things

FluxGym FLUX LoRA Training: Colab, RunPod, Hidden Settings

Guide to FluxGym FLUX LoRA training on Colab and RunPod — including the hidden Kohya settings every trainer should know.

Training LoRAs

Flux LoRA Training Tools Compared

Compare Civit Trainer, Kohya_ss, ai-toolkit, and fluxgym for training Flux LoRA models, with setup notes and source links.

Training LoRAs

FLUX.2 Klein Face Swap: Workflow, Color Matching, Failures

How to swap faces with FLUX.2 Klein using LanPaint think mode and Lab color transfer — plus where the workflow breaks down.

Fixing things

How LoRAs Work and How to Use Them in Stable Diffusion

Understand LoRA mechanics: weight changes at inference, strength controls, stacking multiple adapters, matching base models, and prope…

Training LoRAs

HunyuanVideo img2vid LoRA: Settings & LTX-Video Comparison

Run HunyuanVideo img2vid with the Leapfusion LoRA, compare against LTX-Video, and tune denoise and attention for stable frames.

AI video

Local Video Models Comparison

Which local video models actually run on consumer GPUs? Mochi, LTX-Video, CogVideoX, and Wan2.1 compared for RTX 4090+ hardware.

AI video

LoRA Training Settings: What Works and What Doesn't

Which LoRA settings actually work, and how to test them yourself.

Training LoRAs

Finding and training male character LoRAs

Why male character LoRAs feel scarce on Civitai, what the verified training metadata and directories actually show.

Training LoRAs

Merging LoRAs vs Stacking

When to merge LoRAs into one adapter vs stacking them, the weight math, and the quality trade-offs.

Training LoRAs

Stable Diffusion Model Merging: SuperMerger vs Chattiori

Compare SuperMerger and Chattiori for Stable Diffusion merges. Block-weighted merging, add difference, and cosine similarity evaluatio…

Fixing things

Neta Lumina Prompting: System Tags and Model Comparisons

How Neta Lumina parses prompts, its @artist and #character tag format, and how it compares to Illustrious XL and Pony Diffusion.

Stable Diffusion Consistent Character Without LoRA

Prompt techniques for consistent SD characters without LoRA training.

Training LoRAs

Prompt Vocabulary Cheat Sheet: Photography Terms That Work

Verified photography and cinematography terms for AI image generation prompts, with methodology for building your own cheat sheet from…

A1111/Forge Prompt Tooling: Wildcards and Dynamic Prompts

Compare extensions that clean prompts, generate variations, and manage character libraries in A1111 and Forge — with verified features…

Prompting Fundamentals That Transfer Across Models

What the text encoder receives, why token order matters, and which prompting habits survive SD, SDXL, and Flux.

Stable diffusion empty prompt inpainting comparison

Compare A1111 vs ComfyUI for inpainting with empty prompts and when prompts hurt image edits

Qwen-Image in ComfyUI: Offset Fix and Setup

Qwen-Image setup in ComfyUI: model files, text encoder config, the image-offset problem solution, and comparison with FLUX and SDXL.

ComfyUI

LoRA Regularization Images: When to Use Them

Learn when regularization images help vs. waste compute in LoRA training, how to source them, and the kohya_ss folder conventions.

Training LoRAs

SDXL LoRA Training: Colab vs Hosted Trainers

Compare free T4 GPUs on Colab vs Kaggle for SDXL LoRA training, understand session timeouts, and discover hosted trainer alternatives.

Training LoRAs

SDXL LoRA Training Guide: Lessons from 10 Released Models

Expert SDXL LoRA training guide covering dataset choice, caption style, and hyperparameters from 10 photorealistic model creators.

Training LoRAs

Fast SDXL: Turbo, Lightning & Hybrid Checkpoints

SDXL Turbo and Lightning use 1–8 steps with CFG=0. Here is what each distillation method imposes and what you trade away.

Fixing things

Sketch to Image with Scribble and Lineart ControlNet

How much detail your sketch needs, which ControlNet type to use, and the denoise budget that preserves your composition through upscal…

ComfyUI

SUPIR restoration guided sampling on 12GB GPUs

How SUPIR restoration-guided models rebuild degraded detail, with SeedVR2 video workflows

Fixing things

Synthetic Faces: Training Sets and the Diversity Trap

How AI-generated portrait datasets enable consistent character training while avoiding real-person likeness issues — and the diversity…

Training LoRAs

Train LoRAs on Limited Hardware

Free Colab paths, AMD ROCm setups, and VRAM guidance for LoRA training.

Training LoRAs

AnimateDiff Video-to-Video: Flicker & ControlNet

Prepare source clips, suppress frame flicker, and assemble the ControlNet stack for readable motion in AnimateDiff video-to-video gene…

AI video

AI Video Upscaling and Interpolation Order of Operations

Learn the correct pipeline order for upscaling, interpolating, and finishing AI-generated video without flicker or artifacts.

AI video

Wildcards & Dynamic Prompts: A1111, Forge, Fooocus

Wildcard syntax and dynamic prompt generation in A1111, Forge, and Fooocus

Cut VRAM Use on Windows: HAGS, Page File, Overhead

What actually frees VRAM for AI workloads on Windows — measured settings, documented mechanisms, and community estimates without indep…

Fixing things

Z-Image Turbo Quality: DyPE & Resolution Optimization

Master Z-Image Turbo: DyPE sampling, optimal resolution, and step budgets for maximum quality

Models of 2026

Anima LoRA Training: Settings, Samplers, and Local Setup

How to configure Anima for image generation, pick the right sampler, and train a LoRA locally with verified settings and hyperparamete…

Training LoRAs

Captioning Tools 2026: JoyCaption vs WD Tagger vs Manual

Compare JoyCaption Beta One's 12 caption styles and word-count control against WD Tagger's ONNX batch inference, plus manual tagging w…

Training LoRAs

Cinematic Film Look: Film Stock and Lens Vocabulary for AI

How to use film stock names, lens terms, lighting vocabulary, and grain techniques in AI prompts — and where the prompt stops working.

RunPod vs Vast.ai for LoRA Training on Rented GPUs

Choose the right cloud GPU instance, compare templates, and keep costs low.

Training LoRAs

ComfyUI Acceleration Stack Version Matrix

Version matrix for Triton, SageAttention, FlashAttention, and xformers with PyTorch and CUDA builds for ComfyUI on Windows.

ComfyUI

ComfyUI for A1111 Users: Graph, Nodes, Shortcuts, Habits

Node graph guide for A1111 migrants: read the graph, five starters, keyboard shortcuts, habits that prevent broken workflows.

ComfyUI

Modular ComfyUI: One Graph for txt2img, img2img, Upscale

Build a single ComfyUI graph for txt2img, img2img, and upscale using bypass, KSampler denoise, and reusable templates.

ComfyUI

ComfyUI Upscale + FaceDetailer Chain: Node Order & Settings

Correct node order for upscale then FaceDetailer, and guide_size/force_inpaint settings that prevent full-frame repaints.

ComfyUI

Create an Inpainting Checkpoint from Any Model

Merge SD checkpoints into inpainting variants with Add Difference. Why inpaint weights differ and how to control seed and noise for re…

Fixing things

Build a LoRA Dataset: Sources, Cleanup, Captioning Tools

A practical guide to sourcing images, cleaning data, and using JoyCaption and WD14 tagger for LoRA training dataset preparation.

Training LoRAs

Fix AI Hands: Why They Fail and the Real ControlNet Workflow

Viral Civitai hand-fix embeddings don't exist. Real workflow: ControlNet OpenPose detection + targeted inpainting.

Fixing things

FLUX.2 Klein 9B Local Deploy, LoRAs vs Flux.1-dev

A guide to running FLUX.2 Klein 9B locally, using consistency LoRAs for character work, and comparing it with Flux.1-dev

Models of 2026

Flux ComfyUI Workflows: Base, GGUF, SDXL Refiner, 360°

Flux workflows in ComfyUI — base text-to-image, GGUF for low VRAM, SDXL refiner for animals, equirectangular 360° panoramas.

ComfyUI

Leaving A1111 in 2026: Forge, SwarmUI, SD.Next Compared

Forge, SwarmUI, and SD.Next as A1111 replacements — what breaks, what survives, extension migration guide.

Fixing things

Train a HunyuanVideo Character LoRA

Dataset prep, captioning, and diffusion-pipe settings for HunyuanVideo character LoRA training with VRAM requirements.

AI video

HunyuanVideo: FP8 vs BF16, ComfyUI Setup, LoRA Training

Pick the right HunyuanVideo model file for your GPU, configure ComfyUI, and train a video LoRA.

AI video

Illustrious & NoobAI: Characters, Prompting, LoRA Training

Which anime characters Illustrious XL and NoobAI know natively, how to prompt them with Danbooru tags, and practical LoRA training set…

Models of 2026

Inpainting & Outpainting: Masks, Denoise, Model Choice

When to use inpainting models vs base checkpoints, how masked content and inpaint area change results, and why edits fail to blend — w…

Fixing things

KREA 2 Turbo: Photorealism Settings & Prompt Framing Guide

Verified KREA 2 Turbo photorealism settings, prompt framing for zoom control, and 2026 model comparisons.

Models of 2026

LoRA Library: Folders, Previews, ComfyUI Manager & Kohya

Organize LoRA models with folder conventions, preview naming, and metadata. Set up ComfyUI Lora Manager and Kohya GUI.

Training LoRAs

Fixed-Seed Checkpoint and LoRA Comparison

A practical methodology for fair Stable Diffusion model comparisons — control variables and read grids honestly.

Training LoRAs

Photorealistic Upscaling: Frequency Separation & Workflows

A practical guide to two-stage upscaling, frequency separation for relighting, Ultra-HD prompts, and upscaler comparison.

Fixing things

Pony & Illustrious Prompt Syntax: Score, Source, Rating Tags

Pony V6/V7 and Illustrious XL prompting: score_ quality tags, source_ origin tags, rating_ filters, weights, BREAK, and negative promp…

Models of 2026

Pony Diffusion XL: Native Tags, Testing & LoRA

Which characters Pony XL knows natively, how to test baked-in tags, and when a LoRA is still required.

Models of 2026

SDXL Prompt Structure for Realistic Photography

Structure SDXL prompts with dual-encoder conditioning, templates, artist tags, and ComfyUI presets for photorealistic results.

Stable Diffusion Model Folders: A1111, Forge & ComfyUI

Where checkpoints, LoRAs, VAEs, embeddings, ControlNets and upscalers go across AUTOMATIC1111, SD.Next/Forge, and ComfyUI — and how to…

Fixing things

Textual Inversion Training Guide 2025: Embeddings vs LoRAs

When to use Textual Inversion over LoRA, Kohya_ss and A1111 settings, and the learning-rate gap explained.

Training LoRAs

Latent vs Model Upscaling: Complete AI Image Guide

Compare latent upscaling, model upscalers, hires fix, and tiled workflows. Learn which method works best for photorealistic vs illustr…

Fixing things

Faster Local Video with SageAttention & TeaCache

Compare five acceleration techniques for ComfyUI video workflows — benchmarks, trade-offs, and install steps for each method.

AI video

WAN 2.1 vs 2.2: ComfyUI T2V, I2V, CausVid, VACE Guide

Choose the right WAN workflow for your clip — native T2V/I2V graphs, CausVid, and VACE editing. Model files and VRAM needs included.

AI video

Running WAN Video on 8-12 GB: GGUF Quantization Guide

Master GGUF quantization, block swapping, and SageAttention for WAN 2.2 on 8-12 GB GPUs. Complete workflow for 50-series cards.

AI video

Z-Image Turbo: Local Setup & Training

Learn to run Z-Image Turbo locally, build img2img and inpainting workflows in ComfyUI, and train character LoRAs with the ostris train…

Models of 2026

Z-Image Turbo LoRA: Dataset, Precision, Adapter, AI Toolkit

Train Z-Image Turbo LoRAs with AI Toolkit: dataset sizing, fp8 quantization, training adapter, sigmoid timestep, and Flux/SDXL differe…

Training LoRAs

Camera Angle Prompts in Stable Diffusion

Camera angle and shot-type vocabulary for Stable Diffusion and SDXL prompts, with verified resolution, CFG, sampler and weighting sett…

FLUX.1 LoRA Training Parameters: The Settings That Matter

FLUX.1 LoRA training parameters explained: dataset prep, captioning, learning rate, network dim and alpha, and VRAM for 12GB–24GB GPUs.

Training LoRAs

IPAdapter vs ControlNet vs Regional Prompter in ComfyUI

Which conditioning tool to use in ComfyUI for style transfer, composition and character consistency — and the settings that work.

ComfyUI

Merge Stable Diffusion Checkpoints: Methods & Block Weights

Merge Stable Diffusion checkpoints: weighted sum vs add difference, SDXL block weights, merge vs LoRA, and how to evaluate the result.

Fixing things

Negative Prompts by Model: What Actually Works

Negative prompts work on SD 1.5 and SDXL but do nothing on Flux dev, SDXL-Turbo and Pony by default. See what steers each model family.

Pony Diffusion score_9 and score_8_up quality tags explained

What score_9, score_8_up, source_pony, and rating_safe do in Pony Diffusion, how Illustrious masterpiece tags differ, and when quality…

Models of 2026

ComfyUI Import Errors: Fix Python Env & CUDA Mismatches

Resolve ComfyUI import failures, onnxruntime-gpu CUDA mismatches, and embedded Python package location issues with verified fixes.

ComfyUI

Fix Faces and Hands with ADetailer and FaceDetailer

Compare ADetailer for A1111/Forge and FaceDetailer for ComfyUI — automatic inpainting that detects faces, hands, and persons.

Fixing things

Stable Diffusion Samplers and Schedulers Explained

Pick the right sampler for SD1.5, SDXL, and Flux — deterministic vs. stochastic, step count trade-offs, and scheduler settings that ma…

Train Your First LoRA Free: Data, Captions, Settings

Free LoRA training guide: dataset prep, WD14/BLIP captioning, and the hyperparameters that actually matter. Colab + local workflows.

Training LoRAs

PYTORCH_CUDA_ALLOC_CONF and VRAM Optimization

Fix CUDA OOM with PYTORCH_CUDA_ALLOC_CONF, optimize FLUX training with blocks_to_swap, and run Wan2.1 and HunyuanVideo on 12 GB GPUs.

AI video

Wan 2.2 High Noise vs Low Noise LoRA Workflow

Sequential ComfyUI chain: high-noise KSampler feeds low-noise KSampler. Correct LoraLoaderModelOnly setup and training guide.

AI video

ROCm for Stable Diffusion on AMD: Limits, Gaps, VRAM Cuts

ROCm setup guide for Stable Diffusion on AMD: supported GPUs, Windows gaps, and VRAM reduction techniques.

Fixing things

JoyCaption, TagGUI, and AICaptionHelper

Compare three batch captioning tools for Stable Diffusion datasets

Training LoRAs