174 lines
9.6 KiB
Markdown
174 lines
9.6 KiB
Markdown
# Ideogram 4 — author-informed T2I tests
|
||
|
||
This section follows the model card and author workflow: structured Ideogram 4
|
||
JSON captions, AuraFlow shift 5.25, StarSampler, and CFG 1.0. It is intended
|
||
for controlled quality/time comparisons on the RTX 3070 server.
|
||
|
||
## Current direction
|
||
|
||
- Main models: `ideogram4_int8_convrot.safetensors` and
|
||
`Fal_Ideo_4_NSFW-int8_convrot.safetensors`, tested separately
|
||
- Text encoder: `qwen3vl_8b_w4a8.safetensors`, type `ideogram4`
|
||
- VAE: `flux2-vae.safetensors`
|
||
- Prompt: Ideogram 4 structured JSON caption
|
||
- Sampling: AuraFlow shift `5.25`; StarSampler, Euler, simple scheduler
|
||
- Steps: author workflow 8; model-card reference 10
|
||
- CFG: `1.0`
|
||
- Seed: 20261012 for the controlled baseline
|
||
|
||
The Fal model is retained as a comparison only; it produced more stylised,
|
||
noisier results. The hosted `IdeogramV4` API node was not used because this
|
||
server is not authenticated for Comfy API partner nodes.
|
||
|
||
## Author-inspired comparison
|
||
|
||
The model author publishes a workflow at
|
||
[`Winnougan/Ideogram_Instant_NSFW`](https://huggingface.co/Winnougan/Ideogram_Instant_NSFW).
|
||
It suggests a useful comparison branch rather than a replacement for the
|
||
local baseline:
|
||
|
||
- `Ideogram4PromptBuilderKJ` for structured prompt construction
|
||
- `ModelSamplingAuraFlow` with `shift=5.25`
|
||
- `StarSampler` with 8 steps, Euler, simple schedule, and CFG 1.0
|
||
- Fal Ideogram checkpoint with the local `flux2-vae.safetensors`
|
||
|
||
The author's graph references `Boogu\\qwen3vl_8b_int8_convrot.safetensors`, which
|
||
is not installed here, so the comparison uses the available
|
||
`qwen3vl_8b_w4a8.safetensors`. Complete settings are recorded in
|
||
[`author-inspired-settings.json`](author-inspired-settings.json).
|
||
|
||
These settings remain comparison-only until a fixed prompt and seed are tested
|
||
at 768×1024 and reviewed for typography, anatomy, texture, composition, and
|
||
runtime.
|
||
|
||
The model card also states that the fine-tune expects a structured JSON caption
|
||
with `high_level_description` and `compositional_deconstruction` fields, and
|
||
gives 10 steps with guidance scale 1.0 as its reference usage. The published
|
||
ComfyUI graph uses 8 steps, so both settings are worth measuring separately.
|
||
The card contains several unfinished placeholder sections, so its training and
|
||
fine-tuning details should not be treated as verified claims.
|
||
|
||
## Initial observations
|
||
|
||
- The 512×512 local smoke test completed successfully.
|
||
- The 768×1024 standard Ideogram 4 run took about 6 minutes at 12 steps.
|
||
- The 768×1024 Fal comparison took about 7 minutes at 12 steps.
|
||
- The 8-step Fal run completed in about 5 minutes 18 seconds, but remained
|
||
painterly and noisy.
|
||
|
||
## Current smoke tests
|
||
|
||
The first smoke tests use the same 512×512 structured caption and seed before
|
||
full-size tests are queued. Exact prompt IDs are recorded in `smoke-tests.json`.
|
||
|
||
| ID | Model | Steps | Sampler | Scheduler | Shift | CFG | Generation time (s) | Status |
|
||
|---|---|---:|---|---|---:|---:|---:|---|
|
||
| IDEO-SMOKE-FAL-10 | Fal Ideogram | 10 | Euler | simple | 5.25 | 1 | 45.483 | Complete |
|
||
| IDEO-SMOKE-LOCAL-08 | Local Ideogram | 8 | Euler | simple | 5.25 | 1 | 29.268 | Complete |
|
||
|
||
Smoke results: both completed without node errors. Fal took 45.483 seconds and
|
||
was visually strong. Local took 29.268 seconds and was coherent but notably
|
||
underexposed. The next full-size test should therefore start with Fal; local
|
||
lighting/shift changes should be isolated rather than mixed into sampler tests.
|
||
|
||
## Revised quality/time matrix
|
||
|
||
The previous free-form-prompt/dual-guider queue was stopped because its
|
||
outputs did not meet the quality bar. Those IDs remain historical in
|
||
`submissions.json`; they are not part of this new matrix.
|
||
|
||
Keep the prompt and seed fixed for sampler tests. Keep all other settings fixed
|
||
when changing one variable. A row is not promoted until detail, anatomy,
|
||
composition, text fidelity, artefacts, and runtime are reviewed.
|
||
|
||
| ID | Model | Resolution | Steps | Sampler | Scheduler | Shift | CFG | Generation time (s) | Purpose | Status |
|
||
|---|---|---:|---:|---|---|---:|---:|---:|---|---|
|
||
| IDEO-AUTHOR-01 | Fal | 768×1024 | 8 | Euler | simple | 5.25 | 1 | — | Author sampling | Complete |
|
||
| IDEO-AUTHOR-02 | Fal | 768×1024 | 8 | Euler | simple | 5.25 | 1 | — | Fal author sampling | Complete |
|
||
| IDEO-CARD-10 | Fal | 768×1024 | 10 | Euler | simple | 5.25 | 1 | — | Model-card step count | Complete |
|
||
| IDEO-SAMPLER-01 | Fal | 768×1024 | 8 | res_multistep | simple | 5.25 | 1 | — | Sampler comparison | Complete |
|
||
| IDEO-SAMPLER-02 | Fal | 768×1024 | 8 | euler_ancestral | simple | 5.25 | 1 | — | Sampler comparison | Complete |
|
||
| IDEO-SCHED-01 | Fal | 768×1024 | 8 | Euler | beta | 5.25 | 1 | — | Scheduler comparison | Complete |
|
||
| IDEO-SHIFT-01 | Fal | 768×1024 | 8 | Euler | simple | 3.0 | 1 | — | Shift comparison | Complete |
|
||
| IDEO-SHIFT-02 | Fal | 768×1024 | 8 | Euler | simple | 7.0 | 1 | — | Shift comparison | Complete |
|
||
| IDEO-STEPS-01 | Fal | 768×1024 | 6 | Euler | simple | 5.25 | 1 | — | Speed lower bound | Complete |
|
||
|
||
Quality ranking: `IDEO-CARD-10`, `IDEO-SCHED-01`, `IDEO-SHIFT-02`,
|
||
`IDEO-AUTHOR-02`, `IDEO-AUTHOR-01`, `IDEO-SAMPLER-02`, `IDEO-STEPS-01`,
|
||
`IDEO-SHIFT-01`, `IDEO-SAMPLER-01`. Several lower-ranked outputs contain a
|
||
thin whisker/tendril artefact near the dragon's mouth.
|
||
Results belong in `results/<test-id>.png` with a matching JSON record containing
|
||
the exact prompt, seed, resolution, settings, runtime, and visual verdict.
|
||
|
||
## Reuse
|
||
|
||
Replace `STRUCTURED_JSON_CAPTION` in `workflow-author-inspired.json` with the
|
||
caption from `smoke-tests.json` or a new test caption. Keep the seed and all
|
||
other settings fixed while changing one matrix variable.
|
||
|
||
## Magazine-cover settings matrix
|
||
|
||
The earlier magazine batches were discarded because their prompts and seeds
|
||
varied between runs. The replacement matrix uses one fixed structured magazine
|
||
caption, seed `20261234`, the Fal checkpoint, and the same VAE/text encoder.
|
||
Only the listed technical variable changes per row. Prompt IDs and exact values
|
||
are tracked in [`magazine-cover-matrix.json`](magazine-cover-matrix.json).
|
||
|
||
| Group | Rows | Variable under test |
|
||
|---|---|---|
|
||
| Baseline/steps | 01–03 | 10, 8, and 6 steps |
|
||
| Sampler | 04–05 | `res_multistep`, `euler_ancestral` |
|
||
| Scheduler | 06–07 | `beta`, `sgm_uniform` |
|
||
| CFG | 08–09 | 1.5 and 2.0 versus baseline 1.0 |
|
||
| AuraFlow shift | 10–11 | 3.0 and 7.0 versus baseline 5.25 |
|
||
| Resolution | 12 | 512×768 versus 768×1024 |
|
||
|
||
Expanded rows 13–24 add `euler_cfg_pp`, `heun`, and `dpmpp_2m`; Karras and
|
||
exponential schedulers; CFG 0.8, 1.2, and 3.0; AuraFlow shifts 4.0 and 6.0;
|
||
12 steps; and a 1024×1024 resolution test. They use the same fixed caption
|
||
and seed as rows 01–12.
|
||
|
||
### Magazine matrix review
|
||
|
||
All 24 rows completed. Runtime is wall-clock generation time on the RTX 3070;
|
||
quality scores are a practical review of garment design, model/anatomy,
|
||
composition, typography, and artefacts (10 = strongest overall result).
|
||
|
||
| Rank | ID | Settings summary | Runtime | Quality | Review |
|
||
|---:|---|---|---:|---:|---|
|
||
| 1 | MAG-MATRIX-24 | 1024×1024, Euler/simple, 10 steps, CFG 1.0, shift 5.25 | 30.154s | 8.7 | Strongest garment variation and cleanest overall fashion image; square crop changes the composition. |
|
||
| 2 | MAG-MATRIX-20 | 768×1024, Euler/simple, 10 steps, CFG 3.0 | 152.069s | 8.5 | Excellent asymmetric couture design and pose; slow, with weaker small cover text. |
|
||
| 3 | MAG-MATRIX-09 | 768×1024, Euler/simple, 10 steps, CFG 2.0 | 149.682s | 8.4 | Strong garment structure and silhouette; slow and side text remains synthetic. |
|
||
| 4 | MAG-MATRIX-12 | 512×768, Euler/simple, 10 steps, CFG 1.0 | 20.074s | 8.2 | Good alternate outfit and relatively clean cover hierarchy; lower resolution. |
|
||
| 5 | MAG-MATRIX-05 | 768×1024, Euler-ancestral/simple, 10 steps, CFG 1.0 | 20.865s | 8.1 | Distinct tailored denim outfit and good anatomy; typography is less reliable. |
|
||
| 6 | MAG-MATRIX-08 | 768×1024, Euler/simple, 10 steps, CFG 1.5 | 144.742s | 8.0 | Strong fitted denim design and pose; very slow, with noisy feature text. |
|
||
|
||
The practical production candidate is **MAG-MATRIX-12** for speed/quality,
|
||
with **MAG-MATRIX-24** as the best visual direction if square output is
|
||
acceptable. MAG-MATRIX-16 and MAG-MATRIX-17 are rejected because of severe
|
||
ghosting/blur. The high-CFG runs (08–09 and 20) produced more assertive
|
||
garment structure but unexpectedly long runtimes on this server, so they are
|
||
not default settings without a specific quality need.
|
||
|
||
### MAG-MATRIX-05 double-resolution rerun
|
||
|
||
MAG-MATRIX-05 was rerun with only the resolution changed from 768×1024 to
|
||
1536×2048. The same prompt, seed `20261234`, Fal checkpoint, Euler ancestral
|
||
sampler, simple scheduler, 10 steps, CFG 1.0, and AuraFlow shift 5.25 were
|
||
retained. It completed successfully in **201.965s**. The result has stronger
|
||
denim construction and detail, but the roughly 10× runtime makes it a quality
|
||
mode rather than the default fast workflow.
|
||
|
||
### Research magazine batch
|
||
|
||
The next production batch contains **150 covers** based on the 22 topic
|
||
folders in the local Research repository. It uses the successful double-
|
||
resolution baseline: Fal Ideogram checkpoint, 1536×2048, 10 steps, Euler
|
||
ancestral, simple scheduler, CFG 1.0, AuraFlow shift 5.25, and the same VAE
|
||
and text encoder. Seeds increment from `20261235` so the covers are distinct.
|
||
Each prompt requires English-only, correctly spelled, legible typography and
|
||
neat placement. Named people are treated as adult celebrity subjects in an
|
||
editorial photoshoot; disputed material is framed as claims, reports, or
|
||
documented history in the cover language. Submission IDs and topic mapping
|
||
are recorded in `research-magazine-150-submissions.json`.
|