═══════════════════════════════════════════════════════════════════════════════ DISCO ELYSIUM ═══════════════════════════════════════════════════════════════════════════════ ─────────────────────────────────────────────────────────────────────────────── IDENTITY ─────────────────────────────────────────────────────────────────────────────── Base model Ideogram 4 Generation mode not declared by the author Checkpoint not determined — verify before use Category Visual style Author wystanbelov785 Source https://civitai.com/models/2487755 Provenance CivitAI · model 2487755 · version 3129598 ─────────────────────────────────────────────────────────────────────────────── WEIGHTS ─────────────────────────────────────────────────────────────────────────────── File MixedElysium_Ideogram.safetensors Version Ideogram Size 201.9 MB Format SafeTensor / unknown precision SHA256 e762d7696337f6e50f78a7316a37ef2bcea847ab9a712f5ac5f1fa91697c0caa Published unknown ─────────────────────────────────────────────────────────────────────────────── TRIGGER WORDS ─────────────────────────────────────────────────────────────────────────────── No trigger word. The LoRA applies as soon as it is loaded — strength is the only way to dial the effect up or down. ─────────────────────────────────────────────────────────────────────────────── RECOMMENDED SETTINGS ─────────────────────────────────────────────────────────────────────────────── The author did not publish specific settings. No author settings were published. Community baseline: steps 48 is the reference default: docs/inference.md says "`V4_QUALITY_48` is the default" and the pipeline's `num_steps` default is 48; diffusers documents the same 48-step default. 20 is the practical ComfyUI default — the `V4_DEFAULT_20` preset and the value the shipped template `image_ideogram4_t2i.json` selects (its preset combo defaults to "Default"). 12 = `V4_TURBO_12`. With the TurboTime LoRA: 8 in the nomadoor community workflow; the model card claims usable output down to 2., cfg A guidance SCHEDULE, not one number, and the negative branch comes from a second model rather than a negative prompt. Verified from src/ideogram4/sampler_configs.py: V4_QUALITY_48 = 45 steps @ gw 7 then 3 polish steps @ gw 3; V4_DEFAULT_20 = 18 @ 7 + 2 @ 3; V4_TURBO_12 = 11 @ 7 + 1 @ 3. Pipeline default `guidance_scale` = 7.0 when no schedule is given. In ComfyUI the shipped template sets `DualModelGuider` cfg 7 plus a `CFGOverride` node with widgets [3, 0.7, 1] (cfg 3 from 0.7→1.0 of the run); kijai recommends starting the override at 0.7, not 0.9. TurboTime: cfg 1.0 / no CFG., sampler Euler flow-matching. In ComfyUI: `KSamplerSelect` → euler, driven by `SamplerCustomAdvanced`, with the model-specific `Ideogram4Scheduler` (logit-normal sigma schedule). Its params are `mu` and `std`, not the usual `shift`. Reference presets: mu 0.0 / std 1.5 (48-step), mu 0.0 / std 1.75 (20-step), mu 0.5 / std 1.75 (12-step). The Ideogram4Scheduler node's own widget defaults are steps 20, mu 0.5, std 1.75, but the shipped template overrides mu/std/steps from the preset JSON, so out of the box it runs 20 / mu 0.0 / std 1.75. mu is auto- shifted by resolution in both implementations: mean = mu + 0.5*log((w*h)/(512*512))., resolution 1024×1024 default (pipeline `height`/`width` default 1024). Any multiple of 16 per side from 256 to 2048, aspect ratios up to 6:1 or 1:6 — one model covers all of it. 2048×2048 with V4_QUALITY_48 for maximum quality. Buckets documented in docs/inference.md: 1536×1024 (3:2), 1024×1536 (2:3), 1920×1088 (~16:9), 2048×768 (~21:9), 1024×1792 (~9:16), 1600×400 (4:1)., lora_strength No single official number; these are author-stated values. DeverStyle Archer: 0.6, and explicitly "on the model only, not on the model_negative". zjourney v2: 0.55 standalone (range 0.35–0.6), or v1 @ 0.35 + v2 @ 0.45 stacked. Furry Enhancer: the model page says 0.6–0.8, but its own version notes ask for 0.7–0.9 (v1.94) and 0.8–0.95 (v2.8). TurboTime: 0.6 in the nomadoor workflow (the author gives no figure). So ~0.5–0.9 is the realistic band, read the specific card.. ─────────────────────────────────────────────────────────────────────────────── ADOPTION & POPULARITY ─────────────────────────────────────────────────────────────────────────────── No statistics available. ─────────────────────────────────────────────────────────────────────────────── COMMUNITY NOTES ─────────────────────────────────────────────────────────────────────────────── Archive summary: Disco Elysium is a style LoRA designed for Ideogram 4.0 checkpoints, published by [wystanbelov785](/users/wystanbelov785) on CivitAI in March 2026, now on version 3. Load a compatible checkpoint first, then apply this LoRA on top via your UI's LoRA loader. It has attracted 37 downloads. It's intended for digital art, oil painting, illustration, and cartoon content. The model is flagged NSFW on CivitAI and can produce explicit content. Source: https://civarchive.com/models/2487755?modelVersionId=3129598 Comments are other users' words, reproduced as written. ─────────────────────────────────────────────────────────────────────────────── CHAINING / STACKING ─────────────────────────────────────────────────────────────────────────────── Stacks with others not stated The author gave no chaining guidance. The rules below are the general Ideogram 4 ones, not anything specific to this LoRA. General Ideogram 4 rules that apply here (see REFERENCE.txt): • Chain `LoraLoaderModelOnly` nodes in series off the CONDITIONAL model loader, ahead of the guider. DeverStyle states its samples were made "by setting the LORA @ 0.6 strength on the model only, not on the model_negative". • Model-only loaders throughout. The text encoder is a frozen Qwen3-VL-8B with no LoRA-patchable CLIP branch, so the model+clip `LoraLoader` has nothing useful to do. • To affect the unconditional/negative branch you must load a second, physically distinct model file and patch that one. Passing the same patched model object into both DualModelGuider inputs is a verified no- op (Comfy-Org issue #14521; kijai confirmed it is expected behaviour). • Trigger words go INSIDE the JSON caption, in the semantically matching field, not prepended to a prose string. DeverStyle instructs putting `dvr_arch_tv` in `style_description.art_style` (and in the content field); the Kiki LoRA card describes using `art_style: "Studio Ghibli style"` with `medium: "anime_screencap"`. • Two versions of the same LoRA can be stacked at reduced strengths: zjourney's author recommends v1 @ 0.35 + v2 @ 0.45 over either alone at full strength. • TurboTime changes the graph, not just the strength: it runs at cfg 1.0 and the reference community workflow swaps `DualModelGuider` for a single-model `CFGGuider`, dropping the unconditional model. Re-tune any stacked style LoRA downward. • A LoRA can be applied to only part of the run. The Gray Screen bypass LoRA is used at NEGATIVE weight (about -0.25) and, in its shipped workflows, only on the first step and then disabled; its author warns that "LoRA being active through the whole generation could lead to degradation". • The author never states which weight family this targets. Silent degradation, not hard errors, in every case observed. (1) Loading a conditional-trained LoRA onto the unconditional transformer succeeds silently because the architectures match; it perturbs the negative branch instead of the positive one. (2) Verified ComfyUI trap: in `Du ─────────────────────────────────────────────────────────────────────────────── PROS ─────────────────────────────────────────────────────────────────────────────── + Compact at 202 MB — cheap to stack with other LoRAs. [source: file size] ─────────────────────────────────────────────────────────────────────────────── CONS & CAVEATS ─────────────────────────────────────────────────────────────────────────────── - No trigger word — the LoRA is always on once loaded, so strength is the only control. [source: model metadata] - Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description] - Thin documentation — expect trial and error. [source: description length] - Modest adoption (251 downloads) — less community feedback to rely on. [source: source stats] ─────────────────────────────────────────────────────────────────────────────── COMPATIBILITY WARNING ─────────────────────────────────────────────────────────────────────────────── Silent degradation, not hard errors, in every case observed. (1) Loading a conditional-trained LoRA onto the unconditional transformer succeeds silently because the architectures match; it perturbs the negative branch instead of the positive one. (2) Verified ComfyUI trap: in `DualModelGuider`, passing the SAME model object as both the conditional model and the LoRA-patched unconditional model is a no-op. Comfy-Org issue #14521 tabulates it — case 3 (DualModelGuider, conditional model as positive, conditional model + LoRA as negative) was "exactly identical" to plain conditional-only CFG, while case 4 with the real unconditional model differed. kijai confirmed it is expected: "there's no (working) mechanic to swap LoRA on each step on the same model, it would only work if the model was a different one." To patch the negative branch you must load a genuinely separate model file. (3) A diffusers/PEFT-prefixed LoRA in ComfyUI is expected to produce only a "lora key not loaded" console warning and no visible effect — this follows from the absent key mapping (see checkpoint_families) rather than from a published Ideogram-4 bug report. (4) Cross-quantization loading (fp8-trained LoRA onto nvfp4/nf4/GGUF) generally works with some precision loss; int8_convrot had a specific, now- fixed bug (see gotchas). ─────────────────────────────────────────────────────────────────────────────── AUTHOR'S DESCRIPTION ─────────────────────────────────────────────────────────────────────────────── (the author left no description) ─────────────────────────────────────────────────────────────────────────────── EXAMPLE MEDIA ─────────────────────────────────────────────────────────────────────────────── 6 preview(s) in ./example_images/ — community results for this LoRA. Videos keep a matching _poster.jpg still frame. No prompts are recorded below: CivitAI's public API flags these posts as having generation metadata but does not return it, so the prompts cannot be scraped. Open the model page to read them. 01_image.jpeg (image, 768x1376) 02_image.jpeg (image, 768x1376) 03_image.jpeg (image, 1024x1024) 04_image.jpeg (image, 1024x1024) 05_image.jpeg (image, 1024x1024) 06_image.jpeg (image, 1024x1024) ─────────────────────────────────────────────────────────────────────────────── HOW TO USE (COMFYUI) ─────────────────────────────────────────────────────────────────────────────── Use ComfyUI's NATIVE Ideogram 4 support (Comfy-Org/ComfyUI PR #14259 "feat: Add Ideogram4 model support", merged 2026-06-03) — not the official `ComfyUI-Ideogram4` custom-node wrapper, whose nodes (Magic Prompt / Pipeline Loader / Generate) wrap Ideogram's reference pipeline and contain no LoRA input at all (verified: no occurrence of "lora" in its nodes.py). 1. Update ComfyUI to a build containing `comfy_extras/nodes_ideogram4.py` (which supplies `Ideogram4Scheduler`); the same PR added the generic `DualModelGuider` and `CFGOverride` nodes, which live in `comfy_extras/nodes_custom_sampler.py`. 2. From `Comfy-Org/Ideogram-4` download into `ComfyUI/models/diffusion_models/`: `ideogram4_fp8_scaled.safetensors` (9.28 GB) AND `ideogram4_unconditional_fp8_scaled.safetensors` (9.28 GB). Lower-VRAM alternatives in the same folder: `*_nvfp4_mixed.safetensors` (5.49 GB each) or `*_int8_convrot.safetensors` (9.58 GB each). 3. Download `qwen3vl_8b_fp8_scaled.safetensors` (10.59 GB; or `qwen3vl_8b_nvfp4.safetensors`, 6.31 GB) into `ComfyUI/models/text_encoders/`. The shipped template also lists an optional `gemma4_e4b_it_fp8_scaled.safetensors` from `Comfy-Org/gemma-4` for local prompt building. 4. Download `flux2-vae.safetensors` (0.34 GB, from Comfy-Org/Ideogram-4 or Comfy-Org/flux2-dev) into `ComfyUI/models/vae/`. 5. Put LoRA .safetensors in `ComfyUI/models/loras/` (no subfolder convention enforced). 6. Load the built-in template `image_ideogram4_t2i` (Workflow → Browse Templates), or the workflow from docs.comfy.org/tutorials/image/ideogram/ideogram-v4. 7. Insert a `LoraLoaderModelOnly` between the main `UNETLoader` and the `DualModelGuider`'s conditional model input. There is no CLIP-side LoRA here — the text encoder is a frozen Qwen3-VL — so model-only is the right loader. Strength ~0.5–0.8. 8. Do NOT reuse one patched model object for both guider inputs; it silently does nothing. Patching the negative branch requires a second, physically separate model file with its own loader. 9. Write the prompt as a JSON caption into `CLIPTextEncode` (the CLIPLoader type is `ideogram4`); keep `ConditioningZeroOut` feeding the unconditional branch. Non-ComfyUI paths: diffusers — `Ideogram4Pipeline.from_pretrained("ideogram-ai/ideogram-4-nf4")` (the docs' own example; `ideogram-ai/ideogram-4-nf4-diffusers` also exists), then `pipeline.load_lora_weights(...)`. Training — `ostris/ai-toolkit` (built-in `extensions_built_in/diffusion_models/ideogram4/` plus an `Ideogram4Captioner` that writes JSON captions) or diffusers `examples/dreambooth/train_dreambooth_lora_ideogram4.py`. The official HF weights are GATED (`gated: auto`): accept the license on the model page and authenticate (`hf auth login` / `HF_TOKEN`) or downloads fail. ─────────────────────────────────────────────────────────────────────────────── Scraped by lorakit on 2026-09-05 03:47 UTC. Stats and description are the author's; pros/cons are derived from the evidence tagged beside each line. ───────────────────────────────────────────────────────────────────────────────